ATT Infosoft — Internal PoC
Project Atlas
An AI agent system built to handle IT tender proposals end-to-end. Starting with technical specification writing, Atlas will be trained with company knowledge and purpose-built skills to produce competitive, high-quality tender submissions.
Agent
Atlas
Primary orchestrator
Phase
Phase 04 in progress
First Milestone
TechSpec
Tender tech spec skill
GPU
RTX 5090
32GB VRAM
Objectives
Immediate goal
Train Atlas to read an RFP and produce a structured technical specification section. Evaluate against past ATT submissions and iterate via a feedback loop.
Long-term vision
Atlas dispatches work to specialist agents based on skill match. Each agent holds a role and can carry multiple skills. Output is a complete, review-ready proposal assembled from the right people for each section.
Hardware
| Processor | AMD Ryzen 9 9950X |
| Graphics Card | NVIDIA RTX 5090 32GB — Gigabyte |
| Memory | 64GB DDR5 6400MHz — KingBank RGB |
| Storage | 2TB Kioxia Gen.4 NVMe SSD |
| Motherboard | Asus TUF Gaming B650-E Wi-Fi 6E |
| Power Supply | Seasonic Vertex PX 1200W 80+ Platinum |
| Cooler | Corsair iCue Link Titan RX 360mm AIO |
| Operating System | Windows 11 Pro 64-bit |
| Network | Personal — off ATT infrastructure |
System Design
Architecture
Agents have roles, skills are separate capabilities. Atlas reads the RFP, identifies what's needed, and dispatches each section to the agent with the right skill. Model routing is cost-optimised across three tiers.
System Diagram
Agent Stack
Atlas — Main Agent (Orchestrator)
Powered by OpenAI Codex · GPT-5.4. Handles all user-facing chat, reads RFPs, assigns sections to sub-agents, and assembles the final tender output.
Codex Plus · $30/mo
Qwen3 32B — Primary Local Model
Running on RTX 5090 via Ollama. Q4_K_M quantization. Handles coding, analysis, document processing, tool calling, and tender drafting at zero compute cost.
Free · 24/7
Qwen2.5 Coder 32B — Code Specialist
Secondary local model for code generation, refactoring, debugging, and multi-language script automation. Activated by Atlas for code-heavy tasks.
Free · Local
Qwen3 8B — Fast Triage
Lightweight model for quick routing decisions and simple utility tasks. ~213 tok/s on RTX 5090. Minimal GPU load.
Free · Local
OpenRouter — Fallback Layer
Access to Claude, GPT-4o, and other models when a specific capability is needed beyond Codex or local LLM. Credit-based, used sparingly.
~$10–20 topup
Model Routing
| Task Type | Routed To | Model | Cost |
|---|---|---|---|
| Chat, coordination, RFP reading | Atlas (Cloud) | Codex · GPT-5.4 | $30/mo |
| Coding, analysis, doc processing | Local LLM (5090) | Qwen3 32B | Free |
| Code generation & refactoring | Local LLM (5090) | Qwen2.5 Coder 32B | Free |
| Quick triage & routing | Local LLM (5090) | Qwen3 8B | Free |
| Model-specific edge cases | OpenRouter | Claude / GPT-4o / others | ~$10–20 credit |
| Skill building, quality review | Claude Max (Gabriel, manual) | Claude Sonnet/Opus | Subscription |
Communication Channels
WhatsApp Group
Command centre. One group with Jackie, Sultan, Gabriel, and Atlas. All coordination, status updates, and decisions go through here. Managed via eSIM Bridge ($4.90/mo).
Individual DMs
Each person (Jackie, Sultan, Gabriel) has a separate DM with Atlas for detailed one-on-one work — deep task discussion, private reviews, and context-rich interactions without group noise.
MS Teams
Future plan only. Not included in current PoC — runs off ATT network on personal hardware. May be scoped in once PoC is validated with Jackie.
Memory Architecture
How Atlas and Claude share memory
Claude — Cowork
Gabriel's planning interface. Claude reads and writes the Monday.com shared doc to stay in sync with what Atlas is doing on 5090.
Claude Chat
Cowork
Claude Code
⇄
Shared Bridge
Monday.com
Private Doc · ID 38858462
Human-readable
Read at every session start
Read at every session start
⇄
Atlas — 5090 / OpenClaw
Atlas's persistent local memory. Updates Monday.com shared memory through scheduled consolidation jobs and meaningful milestone updates — not after every session. Reads it on startup to restore shared context before any conversation begins.
MEMORY.md
AGENTS.md
Daily notes
Supermemory
Storage layers
| Layer | Location | Accessible By | Purpose |
|---|---|---|---|
| Shared bridge | Monday.com private doc (ID 38858462) | Claude + Atlas | Cross-agent memory bridge between Atlas and Claude. Append-only writes via scheduled jobs and meaningful milestones. Full bidirectional auto-sync is the direction, not yet fully built. |
| Long-term memory | MEMORY.md — 5090 | Atlas | Curated decisions, preferences, lessons. Updated after every meaningful task. |
| Daily notes | memory/YYYY-MM-DD.md — 5090 | Atlas | Raw activity log. 2am consolidation summarises into MEMORY.md. |
| Agent rules | AGENTS.md — 5090 | Atlas | Operating instructions: "search memory before acting, write after every task." |
| Semantic search | Supermemory plugin | Atlas | Semantic retrieval across all memory within OpenClaw sessions. |
| Local backup | workspace/backups/ — 5090 | Atlas | Rolling 4-week window, day-by-day snapshots (Mon–Sun). Fast local recovery. Automated at 3am nightly. |
| Cold backup | GitHub — bcchew-art/Atlas | Both | Offsite disaster recovery and full version history. Push is manual / on meaningful milestones — not part of the nightly automated backup. |
Automated daily schedule (SGT)
2am
SGT · daily
Memory Consolidation
Atlas reviews today's activity log, distils key decisions, outcomes, and context shifts into MEMORY.md. Appends a timestamped [Atlas] summary to Monday.com so Claude stays in sync.
3am
SGT · daily
Rolling Local Backup
Runs the rolling workspace backup — 4-week window, day-by-day Mon–Sun folder snapshots. Local backup is fast recovery. GitHub push is manual / on meaningful milestones, not automated nightly.
4am
SGT · daily
Rolling Memory Compaction
Conservative pass — today and yesterday stay raw. Content older than 2 days gets merged and deduplicated into MEMORY.md. Keeps memory lean without destroying recent context.
5am
SGT · Sunday
30-Day Deep Prune
Weekly deep clean targeting content older than 30 days. Removes stale or superseded entries. Enduring decisions, preferences, and lessons are always kept regardless of age.
8am
SGT · daily
Morning Briefing
Atlas recalls yesterday's memory and sends a concise morning update to WhatsApp — open tasks, decisions made, what's next. The team starts the day already in context. No catch-up needed.
Feedback Loop
| Input | Past RFP and ATT submission uploaded as reference |
| Evaluate | Atlas compares submission vs RFP requirements, identifies gaps |
| Draft | Generates improved version based on evaluation |
| Review | PM reviews output and provides corrections |
| Store | Corrections saved to skill memory for next run |
Execution
Setup Plan
Four phases from hardware bring-up to first working skill. Currently in Phase 04.
01
Hardware Setup
- ✓Unbox and assemble PCAsus A32 case, Corsair AIO, RTX 5090, Kioxia 2TB SSD
- ✓First boot and BIOS checkEXPO enabled — RAM confirmed at 6400MHz. Kioxia 2TB detected. Mobo: Asus TUF B650-E BIOS 3827
- —Activate Windows 11 ProSkipped for now — no key bundled. Will activate when needed.
- ✓Run Windows Update fully
- ✓Install NVIDIA drivers (RTX 5090 — Blackwell)Driver 595.79 — installed via NVIDIA App
- ✓Verify GPU detected and CUDA enabledRTX 5090 confirmed in Device Manager
02
Core Stack Installation
- ✓Install Ollama v0.17.7 (Windows native)Windows native install, no WSL2
- ✓Pull Qwen3 32BPrimary local fallback — analysis, doc processing, tender drafting
- ✓Pull Qwen2.5 Coder 32BCode specialist — generation, refactoring, debugging
- ✓Pull Qwen3 8BFast triage fallback — ~213 tok/s on RTX 5090
- ✓Install OpenClaw 2026.3.11Windows native + Git for Windows dependency
- ✓OpenClaw onboardingGPT-5.4 primary, WhatsApp allowlist, Brave search API
- ✓Connect Codex OAuthCodex Plus — GPT-5.4
- ✓OpenClaw dashboardFixed — use `openclaw dashboard` to launch with auth
- ✓Configure model routingGPT-5.4 = human-facing chat, orchestration, final review. Qwen3 32B/8B = heavy repetitive work (doc reading, extraction, drafting, classification). Cost split by role, not just speed.
- ✓Install gateway serviceVia openclaw doctor --fix — auto-starts on boot
- —Connect OpenRouter API keyDeferred — not priority for now
- ✓Configure embedding providerOllama + nomic-embed-text (274MB) — free local embeddings
- ✓Add Jackie + Sultan to WA allowlistJackie (+6583229171), Sultan (+6593663700)
03
Atlas Identity and Channels
- ✓Create IDENTITY.mdName, role, emoji, stack, communication style per stakeholder
- ✓Create SOUL.mdCore behavioral rules, boundaries, proactive behavior, continuity
- ✓Create AGENTS.mdSession startup, memory rules, tender work rules, sub-agents, red lines
- ✓Configure MEMORY.md and memory/ directory on 5090Embedding: Ollama + nomic-embed-text, temporal decay 30-day half-life
- ✓Deregister WhatsApp from Mac MiniDone — number freed
- ✓Link WhatsApp number to Atlas on 5090+6598198569 linked via eSIM Bridge
- ✓Atlas first contact + tone calibrationAtlas alive on WA. Tone calibrated to Gabriel's preference. Preference saved to memory.
- ✓Memory persistence verifiedEnd-to-end: user speaks → Atlas adapts → writes memory file → git commits → persists
- ✓WA group policy configuredgroupPolicy: open, requireMention: false. SOUL.md controls listen-first reply behavior.
- ✓Profile picture generatedAI-generated portrait — team voted on final pick. Set as WA profile.
- ✓Create WhatsApp group"Project Atlas" — Jackie + Sultan + Gabriel + Atlas
- —Set up individual DMs with AtlasDeferred — not needed yet
04
Full Tender Pipeline — All 9 Skills Built
- ✓Upload past tender RFPs as referenceCAG T3 IIDS Refresh RFP uploaded — 04A, 04C, 04F Appendix K
- ✓Build Ethan — Technical Design agentLive on ollama/qwen3:32b. Generates 06 + 07 for CAG-style tenders.
- ✓Build Marcus — Project Management Plan agentLive on ollama/qwen3:32b. Generates sections 08, 09, 10 for CAG-style tenders.
- ✓First clean Ollama run — all 5 CAG docs generatedTest 10: stable generation pipeline confirmed. Bundled deterministic scripts resolved syntax errors.
- ✓Redesign agent/skill model — decouple roles from skillsAgents now have job titles. Skills are independent capabilities. Atlas has Tender Orchestration skill.
- ✓Connect Monday.com — Atlas can read/write shared memoryAPI token configured, monday.js CLI built. Append-only write rule prevents Claude/Atlas collision.
- ✓Automate memory schedule — 5 cron jobs2am consolidation, 3am backup, 4am compaction, Sunday 5am 30-day prune, 8am briefing.
- ✓Build Victor — Pricing / BQ agentSmoke test PASS: 29,900 bytes. ollama/qwen3:32b.
- ✓Build Nadia — Company Profile agentSmoke test PASS: 29,812 bytes. ollama/qwen3:32b.
- ✓Build Grace — Management Summary agentSmoke test PASS: 10,632 bytes. Fixed 3 bugs: __dirname path, PageNumber constructor, footer.
- ✓Build Kevin — Quality Evaluation agentSmoke test PASS: 12,876 bytes. Fixed 2 bugs: __dirname path, WidthType.PERCENTAGE value:'%' → size:N crash.
- ✓Build Dev — Functional Response agentSmoke-ready. table_functional_matrix + table_workflow. ollama/qwen3:32b.
- ✓Fix Atlas WhatsApp ack issue — removed ack-before-work ruleAGENTS.md updated: no ack, go straight to tool calls. Session was ending after every acknowledgement.
- ✓Fix model routing — all 9 agents → ollama/qwen3:32bSmoke tests were burning GPT-5.4 Codex tokens. Fixed: exec generator directly, no spawn needed.
- ✓Fix stale agent roster — purged Greek myth placeholders from AGENTS.md + SKILL.mdDaedalus/Midas/Athena/Hestia/Argus replaced with correct 9-agent roster.
- Improve output depth — content quality review with Jackie
Tender Skills
Skills Library
Skills are independent capabilities — not tied to one agent. Atlas assigns the right skill to the right agent based on the RFP. An agent can hold multiple skills, and skills can be reassigned as the team grows.
Atlas Skills
Tender Orchestration
Reads the RFP, identifies all required submission sections, creates the folder structure, and dispatches each section to the agent with the matching skill. Assembles and validates the final output before delivery.
Owner: Atlas
Sub-Agent Skills
Technical Design
Generates the full technical design section: development approach, system architecture, hosting environment, tech standards, application security, and cloud services. Built to map directly to evaluator scoring criteria.
Owner: Ethan — Technical Engineer
Project Management Plan
Generates project team structure, Gantt chart, QA plan, change control procedure, and risk register. Maps directly to the PMP section of the tender submission.
Owner: Marcus — Project Manager
Compliance Matrix
Generates a paragraph-by-paragraph C / NC / PC compliance statement against every RFP clause — General, Functional, and ICT Security. Flags gaps and provides justifications for non-compliances.
Owner: Hana — Compliance Analyst
O&M, Training & Documentation
Generates the post-implementation support plan: L1/L2 service structure, SLA compliance, hardware and software maintenance strategy, patch management, training programme with course syllabi, and full documentation list. Covers CAG sections 11–13 and equivalent O&M/training sections in other tenders.
Owner: Joel — Operations Engineer
Pricing / BQ
Generates structured cost schedule based on scope, mandays, and rates. References past BQ formats and market benchmarks to produce competitive and defensible pricing.
Owner: Victor — Commercial Analyst
Company Profile
ATT Infosoft background, past project references, and relevant track record. Tailored to the domain of each tender to maximise relevance score with evaluators.
Owner: Nadia — Business Development
Functional Response
Feature-by-feature solution description mapped to every functional requirement. Compliance matrix (Full / Partial / Non-Compliant) plus operational workflow descriptions. Typically the largest section in a tender.
Owner: Dev — Solutions Analyst
Management Summary
High-level overview synthesised from all completed sections. Generated last. Frames ATT's value proposition and key differentiators for the evaluating committee.
Owner: Grace — Executive Writer
Quality Evaluation
Cross-checks every section against RFP requirements and evaluation criteria weightings. Scores submission quality, flags missing items, and highlights inconsistencies before final output.
Owner: Kevin — QA Engineer
Activity
Progress Log
Dated record of planning sessions, decisions, and milestones.
16 Mar 2026
Session — Full 9-agent pipeline complete. All smoke tests PASS.
All 9 tender sub-agents now built and smoke-tested. Victor (Pricing/BQ): 29,900 bytes ✓. Nadia (Company Profile): 29,812 bytes ✓. Grace (Management Summary): 10,632 bytes ✓ — required 3 bug fixes (section-content.json path resolved from __dirname not tenderRoot; PageNumber.CURRENT used correctly inside TextRun children; footer patched). Kevin (Quality Evaluation): 12,876 bytes ✓ — required 2 bug fixes (same __dirname path fix; root cause of slice() crash traced to WidthType.PERCENTAGE receiving value:'15%' string instead of size:15 number). Dev (Functional Response): built and smoke-ready (functional_matrix + workflow tables). Additional hardening: Atlas WhatsApp ack-before-work rule removed entirely (session was ending after every acknowledgement message); all 9 agents confirmed on ollama/qwen3:32b routing; stale Greek myth agent names (Daedalus/Midas/Athena/Hestia/Argus) purged from AGENTS.md and Atlas SKILL.md; team profile pictures added for Victor, Nadia, Grace, Kevin, Dev; site widened to 1080px; team grid changed to 4 columns; endgame bar updated to 12/12. Pipeline is now complete end-to-end — all sections of a CAG-style tender can be generated.
16 Mar 2026
Tender 112 — Sub-agent hardening + all 4 generators smoke-tested
First live run with all 4 sub-agents (Ethan, Marcus, Hana, Joel) failed across the board. Root causes traced to skill/generator contract mismatches: Ethan's section-content.json was still in the old legacy shape; Joel and Hana had no committed starter payloads; Hana (Themis) inferred the wrong script path entirely and never ran the generator. Atlas patched all 4 SKILL.md files — absolute generator paths are now explicit, exact output filenames are specified, and placeholder names are forbidden. Valid typed starter section-content.json files committed for all agents. A controlled deterministic smoke test after patching confirmed all 9 output files generated successfully: 06 Technical Proposal (30 KB), 07 System Design (29 KB), 08–10 Project Plan / Governance / Team (44 KB each), 11–13 O&M Plan / Training / Documentation (29 KB each), Compliance Matrix (29 KB). Note: smoke test used pre-seeded starter content, not a fresh full live sub-agent run — that is the next step.
14 Mar 2026
Session 16 — Architecture redesign, Monday.com live, memory automation
Major architecture rethink: decoupled agents from skills — agents now have job roles (Technical Engineer, Project Manager, etc.) and skills are independent capabilities that can be assigned across agents. Atlas gets her own Tender Orchestration skill. Site fully revamped to reflect new model: team tab shows role + skill tags (green = live, grey = planned), Skills Library tab decoupled from agent names with Owner column, endgame progress tracker added (3/9 skills live). Monday.com integration connected — API token configured, monday.js CLI tool built, append-only write rule enforced to prevent Claude/Atlas memory collision. Automated memory schedule wired up: 2am consolidation, 3am rolling local backup (no auto GitHub push), 4am daily compaction (Atlas's enhancement), Sunday 5am 30-day deep prune (Atlas's enhancement), 8am morning briefing already live. Atlas wrote the PowerShell backup script and added the 4am/5am jobs herself while Gabriel was out.
13–14 Mar 2026
Sessions 9–15 — Ethan & Marcus live on Ollama
Built Marcus (PMP sub-agent) covering CAG sections 08, 09, 10. Resolved a 10-test gauntlet to get both Ethan and Marcus running on ollama/qwen3:32b: fixed model routing (spawn-call level parameter), Ollama auth (agent auth store), exec/sandbox conflicts, extracted-file naming, and JS syntax errors in generated scripts. Breakthrough fix: bundled deterministic generator scripts inside each skill (skills/hermes/scripts/generate-cag.js, skills/chronos/scripts/generate-cag.js) instead of asking the model to write them from scratch. Test 10 was the first clean Ollama pass — all 5 CAG documents generated and verified. Content depth (~1,000–1,800 words per doc) is the remaining focus before Jackie review.
13 Mar 2026
Session 8 — Phase 03 complete, Atlas operational
Atlas alive on WhatsApp — first contact successful. Tone calibrated from corporate to casual colleague (preference saved to Atlas memory, git committed). Memory persistence verified end-to-end. WA group policy configured: groupPolicy "open" + requireMention false — SOUL.md controls listen-first reply behavior. QMD memory attempted but incompatible with Windows (requires /bin/sh), reverted to builtin. Profile picture generated (AI portrait, team voted). WA group "Project Atlas" created with Jackie, Sultan, Gabriel, Atlas. Individual DMs deferred. Phase 03 complete.
13 Mar 2026
Session 7 — Atlas identity installed
Atlas persona files created and deployed to OpenClaw workspace: IDENTITY.md (name, role, stack, comm style), SOUL.md (behavioral rules, boundaries, proactive behavior), AGENTS.md (session startup, memory rules, sub-agents, red lines). Embedding provider configured: Ollama + nomic-embed-text for free local semantic search. WA allowlist updated with Jackie and Sultan. Phase 02 complete. Phase 03 in progress — remaining: WA number linking and group creation.
13 Mar 2026
Session 6 — Core stack operational
Dashboard issue resolved — gateway wasn't running; fixed with gateway.cmd and service install via openclaw doctor --fix. PowerShell execution policy set to RemoteSigned (was blocking CLI). All 3 Ollama models confirmed. Codex OAuth connected. Model routing configured: GPT-5.4 primary → Qwen3 8B → Qwen3 32B as fallbacks. GPT-5.4 kept as primary for stakeholder communication quality. OpenRouter deferred. Remaining: embedding provider for memory semantic search, WA allowlist for Jackie + Sultan.
12 Mar 2026
Session 2 — Architecture finalised
Communication strategy updated: WhatsApp group (Jackie + Sultan + Gabriel + Atlas) as command centre, with individual DMs per person. Discord dropped. Memory architecture decided: Monday.com private doc as shared memory across all Claude surfaces and Atlas. Local MEMORY.md + AGENTS.md + Supermemory plugin on 5090, with 30-day temporal decay. Sub-agents named (Greek mythology): Ethan, Victor, Kevin, Hana, Dev, Marcus, Nadia, Grace. Multi-agent fleet (Titan, Forge) deferred — Atlas PoC first. GitHub (bcchew-art/Atlas) confirmed for code and site assets. Cowork confirmed able to deploy directly to GitHub.
12 Mar 2026
Session 1 — Project Atlas scoped
Defined full project scope, agent architecture, model routing strategy, and skill roadmap. Hardware confirmed as RTX 5090 workstation. Agent named Atlas. WhatsApp as primary channel. First milestone set as TechSpec skill. PoC to run off-network on personal hardware per Jackie's direction. Multi-agent fleet diagram designed.
People
Atlas Team
Agents have roles, not just tasks. Each agent is a specialist — Atlas assigns work based on who has the right skill. Skills are independent modules: one agent can hold many, and new skills can be added over time.
Team
Atlas
Lead Orchestrator
Active
Skills:
Tender Orchestration
Ethan
Technical Engineer
Live
Skills:
Technical Design
Compliance Review
Marcus
Project Manager
Live
Skills:
Project Mgmt Plan
Pricing / BQ
Hana
Compliance Analyst
Live
Skills:
Compliance Matrix
Joel
Operations Engineer
Live
Skills:
O&M Plan
Training Programme
Documentation List
Dev
Solutions Analyst
In Progress
Skills:
Functional Response
Victor
Commercial Analyst
In Progress
Skills:
Pricing / BQ
Nadia
Business Development
In Progress
Skills:
Company Profile
Grace
Executive Writer
In Progress
Skills:
Mgmt Summary
Kevin
QA Engineer
In Progress
Skills:
Quality Evaluation