Tessmark Equity Research Analyst Workbench

Project Description

We need an AI Workbench for our equity research analysts. The workflow we want to replace is the morning Bloomberg-terminal scroll. Each analyst covers 20-100 tickers, currently reads 200+ news items per day, sends 5-10 written notes per week to PMs at our buy-side clients. Persona: senior equity analyst. ~3 hrs/day on news triage today. The job-to-be-done is to find the 5 items in the daily flow that move the thesis, write notes on them, share with the PM team. Not the analyst's judgment - the triage + drafting are the targets. Replaces: Bloomberg-scroll + manual triage + open-a-doc-and-write workflow. REFERENCE OPEN-SOURCE PATTERNS (please review when designing M1): - kxsystems/nvidia-kx-samples/tree/main/ai-model-distillation-for-financial-data Production-shaped Data Flywheel (FastAPI + Celery + Redis + MLflow + Kubernetes + KDB-X). Distills Llama 3.3 teacher into a 1B/3B student via LoRA for 13-category event classification. Output is dual-mode: classification + BUY/SELL/HOLD with rationale (signal_config). Validation dual-axis: F1 + Sharpe + max drawdown + win rate. KDB-X Community Edition is free for commercial use per their README (verify before relying). - nvidia-ai-blueprints/quantitative-portfolio-optimization Mean-CVaR + Efficient Frontier + Rebalancing on cuOpt. Apache 2.0. NOT the primary reference for this build (different persona); included for context only. BUILD POSTURE: we are NOT recreating the flywheel. We consume a classifier (Claude Haiku as student-stand-in for the demo) and add the analyst-facing surfaces. Single backend Postgres + pgvector + Django + Celery + Redis (no KDB-X for now). DAILY TEXTURE - 30 dimensions our M1 architecture should reckon with (named without prejudice; some Hale-style workbench traits may bend; some may not apply): 1. Cadence: hourly news cycle (vs Hale's deal-weeks) 2. Ingest cardinality: many sources, unstructured text, dedup non-trivial 3. Output reversibility: notes go to internal PM team (reputation-reversible vs cold-email public-irreversible) 4. Transparency affordance grade: classifier confidence + top-token attribution + top-K alternatives 5. Relevance time-decay: 2-day-old news is dead 6. Coverage overlap: multiple analysts on same ticker (multi-seat per persona) 7. Source attribution requirement: notes must cite sources 8. Calibration loop visibility: analyst sees how well the system has been calibrating 9. Pipeline recursion: training/distillation sub-pipeline separate from runtime pipeline 10. Cost surface composition: AI cost vs (potentially) GPU infra cost 11. Restricted-list refusals: blackouts on names where firm is underwriting/advising - WORKBENCH MUST REFUSE TO DRAFT 12. MNPI quarantine: NDA'd management-call content cannot enter the ingest pipeline 13. Multi-stage human gate: client-facing notes require compliance-officer sign-off in addition to analyst 14. Vendor-data display restrictions: Bloomberg/FactSet quotes have redistribution rules 15. UI replay log: "what was on screen at 3:42pm April 15?" for audit defense 16. Prior-belief baseline: news value = delta from consensus, not absolute 17. Peer-relative context: compare across comparable-set, not just absolute 18. Forward catalyst calendar: events flow backward (news links to upcoming earnings/FDA/Fed) 19. Cross-asset signal graph: news on equity affects bonds, options, supplier names, sector ETFs 20. Source-quality + adversarial filter: Bloomberg vs Twitter vs WSB - threat model required 21. Latency expectation declared: this is the THESIS layer, not the TRADING layer 22. Conviction taxonomy in output: notes carry structured grades (high/thesis-changing/clarifying/housekeeping) 23. Machine-to-machine handoff: notes carry JSON payload (ticker/direction/conviction/horizon) for trading-desk software 24. Position-aware filtering: does our firm own this ticker (long/short/flat)? 25. Sell-side coverage map: who else covers this name + what did they say recently? 26. Per-ticker note half-life: track our own past output per (analyst x ticker) over time 27. Long-horizon performance attribution: 12-month outcome ledger - did our calls land? 28. Bloomberg keyboard inertia: dense monospace + hotkeys, or adoption stalls 29. Earnings-season burst load: 5-10x normal news flow for 4 weeks/quarter 30. Quote-stream cross-reference: inline charts beside text, or browser-jump

Additional Notes

Tessmark Research is a synthetic Rule 47-audited demo persona for the AI Workbench category N=2 methodology test. Per the dimension-pressure-list, M1 should produce architecture that validates or invalidates Hale-derived workbench traits. NOT a real client engagement.