Multi-Agentic Research Platform
Answers research questions only with claims it can verify — every stage traced, every citation grounded.
infra:python · postgres + pgvector · gemini embeddings · typescript · docker
> open dossier> resuming session
> 4 systems operational
> last deployment: 2026.06.12
> channel open
systems engineer · ai infrastructure
DEPTH.01 · SYS.CORE / 01
Designs backend systems where behavior under pressure matters more than the demo: multi-agent research pipelines, market simulation, conversational data systems. The recurring problem: how intelligence behaves under constraints.
local focus: agent pipelines · evidence-grounded retrievalchannel open
DEPTH.02 · SYS.VAULT / 02
4 systems
Answers research questions only with claims it can verify — every stage traced, every citation grounded.
infra:python · postgres + pgvector · gemini embeddings · typescript · docker
> open dossierLets strategies face real market data without letting simulated capital pretend to be real.
infra:fastapi · react + vite + typescript · mongodb (motor) · yahoo finance api · openai api
> open dossierMakes a heavyweight scientific archive answerable in plain language, from a cached store that fits on one machine.
infra:python · streamlit · plotly · leaflet · cesium · netcdf → parquet · ollama (optional)
> open dossierGenerates and grades interviews with an LLM while keeping auth, sessions, and scoring accountable to a real backend.
infra:fastapi · react 19 · mongodb · gemini 2.0 flash · web speech api · jwt auth
> open dossierDEPTH.03 · SYS.LOGS / 03
6 operations · 3 failure traces
SYS.ATS / 02[migration]info
Accounts, sessions, and simulations migrated from process memory to MongoDB through Motor (async). Process restarts were erasing simulated books mid-run.
> outcome: sessions survive restarts; the in-memory store retained as a zero-setup path for local runs.
SYS.ATS / 02[failure]warning
Copilot routed simulation commands through the model alone. Loose prompts produced malformed simulations; valid-looking output failed downstream.
> outcome: hybrid routing shipped — heuristic parsers carry the command path; the model is enrichment.
experiment closed. the model path is assumed to fail and is routed around, permanently.
SYS.FLOAT / 03[rollback]critical
Interactive queries against the raw ARGO NetCDF archive exceeded the latency ceiling — multi-second stalls per question made exploration unusable.
> outcome: rolled back to an offline NetCDF → Parquet converter; the live application reads only the cache.
conversion is now a precondition, not an optimization. the live path does not exist.
SYS.MARP / 01[architecture]info
Single-prompt research answers could not be audited — the model answered but could not show its work. Rewritten as Planner / Retriever / Writer / Critic / Verifier with typed trace events at every stage.
> outcome: each stage inspectable and replaceable; the critique loop bounded by a hard iteration cap.
SYS.MARP / 01[failure]warning
Multi-step retrieval plans truncate to step one; remaining steps are recorded in the trace but never run. Discovered under multi-question research loads.
> outcome: limit preserved in the execution trace rather than papered over; downstream stages treat thin evidence as a first-class state.
still unresolved — full plan execution pending.
NEXUS[deployment]info
First deployment of the descent engine: the plane transform created a containing block that broke ScrollTrigger's pin fixed-positioning during the Vault traverse.
> outcome: plane transform cleared for the pinned stratum's focus band; micro-drift sacrificed — the pin is worth more than a 0.5% creep.
DEPTH.04 · SYS.LIVE / 04
telemetry sourcesgithub rest apibuild metadata> aggregated · revalidated 60s
> source: vercel build env · stamped at deploy
> source: github rest api · revalidated 60s
DEPTH.05 · TERMINAL
> enter executes · arrow-up recalls · esc releases
DEPTH.06 · NODE: CONTACT
If something here is worth building on, I read every transmission.