I didn’t switch careers. I kept moving one layer deeper.
I did not arrive here from AI. I arrived here from running systems that could not afford to be wrong.
SCROLL TO TRACE THE SYSTEM ↓
BUILDING THE MACHINE
First, make software actually run.
I started by building systems — APIs, databases, Linux, web systems, Git, CI/CD, deployment, production debugging, monitoring.
Before managing products, I learned what it means to make software actually run.
The commit is the unit of truth — every later system of mine is reproducible from source.
Era — Daapapp · Infinite8 · SportMob
BUILDING THE RIGHT MACHINE
The technical system works. Then ask: but should it exist?
From “Can we build it?” to “Should we build it, for whom, and what changes if we do?”
Discovery · hypothesis · experimentation · analytics · metrics · marketplace design · growth · ranking · monetization · vendor systems · operational products — learned as chapters, not as a skill cloud.
“A product without engineering discipline is a sketch; an engineering effort without product discipline is a hobby — the interesting work is at the boundary.”
BUILDING THE SYSTEM THAT BUILDS THE PRODUCT
Why do good people inside good companies still build slowly?
The bottleneck was not always the code. Sometimes it was the organization around the code.
Four functions. No shared flow. Work queues at every border.
Agile coaching · product transformation · team design · cross-functional execution · delivery systems · ownership · operating models · feedback loops · PSM I.
WHEN PRODUCTS BECOME SYSTEMS
At marketplace scale, there is no isolated feature.
Every decision moves through a system.
Tap a domain — watch the decision propagate.
This is where product, engineering and operations began converging: end-to-end tracing, root-cause investigation, data-backed decisions, integration thinking, platform thinking, economic thinking.
SnappMarket · Okala · Alibaba Travels — generalized domain map. No employer internals.
CAN A SYSTEM LEARN WHEN NOT TO ACT?
Prediction is cheap. Trustworthy action is expensive.
A real research runtime on a small server — not notebooks. It had to survive:
- continuous live market feeds
- state persistence
- process restarts
- model inference
- feature generation
- delayed outcomes
- experiment tracking
- paper decisions
- evaluation
- resource constraints
THE MICRO-SERVER DISCIPLINE
The machine got visibly more disciplined, not merely more complicated. The early lesson: historical complexity did not equal edge.
05A — THE LEGACY AUDIT
The first useful result was not a profitable model. It was learning that the old system had been confidently wrong.
HYPOTHESISThe old system was a competent base to build on.
WHAT WE BUILTForensic audit of the legacy database: candle reconstruction (a true 4h candle = exactly 4 one-hour candles), 307 closed trades reconstructed.
WHAT REALITY SAID~86% of trades ended STOP_HIT · ~10.75% win rate · realized PnL materially negative. The old system had been confidently wrong.
WHAT CHANGED NEXTSentinel rebuilt around honest evaluation and reconstruction rules instead of inherited belief.
05B — MORE DATA ≠ MORE INTELLIGENCE
More data was not automatically better data.
HYPOTHESISPooling legacy data into current training improves discrimination.
WHAT WE BUILTIdentical schema, labeler and seed — two arms: current-only vs current + legacy, purged chronological holdout.
WHAT REALITY SAIDCURRENT ONLY AUC 0.6848 vs CURRENT + LEGACY 0.6763. Legacy-only zero-shot transfers at chance. Domain shift dominates.
WHAT CHANGED NEXTRelevance and evaluation validity became first-class checks. More data ≠ more intelligence.
05C — QUANTIVIQ SENTINEL
22 base features + 5 regime features → causal HMM → LightGBM → calibration → VON 0.60 gate → PAPER
Live candidates arrive. The calibrated gate decides.
The first production behavior I trusted was not BUY or SELL. It was “I don’t know.”
16,000+ OUTCOMES MATURED, PAPER-ONLY · THE CHALLENGER STILL DID NOT BEAT THE CHAMPION STRONGLY ENOUGH TO PROMOTE
HYPOTHESISA disciplined pipeline can produce calibrated, abstaining paper decisions on live data.
WHAT WE BUILTLive market → candles → features → regime → model → calibration → confidence → policy → abstain or propose → outcome ledger → research loop. VON 0.60 gate, paper-only.
WHAT REALITY SAID16,000+ outcomes matured. At one point the system processed live candidates and chose ABSTAIN for all of them. The challenger never beat the champion strongly enough to promote.
WHAT CHANGED NEXTPromotion gates formalized; live capital stayed disabled.
05D — THE SCIENTIFIC DISCIPLINE
EARLY
LATER
- persist every decision, including ABSTAIN
- attach delayed outcomes after the fact
- preserve counterfactual evidence where possible
- separate champion from challenger
- freeze the incumbent during an experiment
- retrain challengers slowly, not continuously mutate production
- walk-forward / out-of-sample evaluation
- measure calibration, not just classification accuracy
- measure selective precision and coverage
- evaluate net economics after fees and slippage
- compare against simple baselines
- require promotion gates
- keep live capital disabled until evidence justifies authority
- allow a negative result to terminate an experiment
The models changed. The bigger change was learning how not to fool myself with them.
| HMM / STATE | What market regime are we in? |
| LIGHTGBM / DIRECTION | Is there directional evidence? |
| CALIBRATION | Does 70% confidence actually behave like 70%? |
| HUBER MULTI-OUTPUT | What is the likely move / magnitude under noisy targets? |
| MFE / OPPORTUNITY | How much favorable excursion is realistically available? |
| FUSION / POLICY | Even if models disagree, is there enough evidence to act? |
| UNCERTAINTY GATE | Should the system abstain? |
05E — THE EXPERIMENT GENERATIONS
HYPOTHESISThe old system was a competent base to build on.
WHAT WE BUILTForensic audit of the legacy database: candle reconstruction (a true 4h candle = exactly 4 one-hour candles), 307 closed trades reconstructed.
WHAT REALITY SAID~86% of trades ended STOP_HIT · ~10.75% win rate · realized PnL materially negative. The old system had been confidently wrong.
WHAT CHANGED NEXTSentinel rebuilt around honest evaluation and reconstruction rules instead of inherited belief.
The tests passed. The hypothesis did not.
05F — PHOENIX · AUTONOMY MEETS CONSEQUENCE
Every signal must pass verification, economics, risk and authority before anything moves.
Rust · Go · Python · Solidity · PostgreSQL · NATS JetStream · Prometheus · Docker · GitHub Actions · Arbitrum · Aave
- deterministic fixtures for profitable / non-profitable / unsupported / incomplete / duplicate conditions
- startup blocking when critical dependencies are not real or valid
- immutable / versioned build artifacts
- migration and state checks
- health gates before release
- rollback paths
- manual-only transition to live authority
LIVE_EXECUTION=false was not an inconvenience to hide; it was an explicit safety posture while feed correctness, math parity, execution semantics, provider behavior, economic validity and reconciliation were still being proven.
They are not the same claim. Phoenix proved the first two — and only the first two.
- 1Architecture is not edge.
- 2Live is not the same as ready.
- 3Autonomy requires authority boundaries.
- 4Fail-closed is a product decision.
- 5Economic gates belong before execution.
- 6Reconciliation is part of execution.
- 7Productionization can reveal that the hypothesis itself is weak.
- 8Technical difficulty is not commercial validation.
Phoenix did not teach me how to automate risk. It taught me how much structure must exist before autonomy deserves authority.
HYPOTHESISFail-closed autonomous execution infrastructure can earn the right to act.
WHAT WE BUILTRust/Go/Python/Solidity on Arbitrum: signal → verify → economics → risk → authority → execution → reconciliation. Dual-provider agreement, exact economic gates, single-transaction authority, receipt and balance reconciliation, protected release, deterministic fixtures, health gates, rollback.
WHAT REALITY SAIDAt a mature snapshot the containers were up and monitoring was live — with $0 realized revenue, parts of the control plane disarmed by design, an observer producing no actionable bids, and one deployment succeeding while execution readiness stayed false because a provider was rate-limited. 39,538 shadow decisions. 0 executions. Retired as a disciplined shutdown.
WHAT CHANGED NEXTArchitecture is not edge. Live ≠ ready. Autonomy requires authority boundaries. Production discipline was built before permission to risk capital — and that discipline is the reusable asset.
INTELLIGENCE BECAME INFRASTRUCTURE
The model is not the product.
The router ranks providers by learned reliability, not by config.
- OpenAI-compatible endpoint
- multiple providers
- automatic model discovery
- health checks
- capability-aware ranking
- runtime statistics
- Wilson lower bound / reliability scoring
- time-decay
- exploration
- cooldowns
- bounded fallback
- streaming
- local models
- DSH integration
- privacy-first / zero telemetry
The model is not the product. Reliable intelligence requires routing, economics and failure handling.
FROM MODEL TO WORKER
A capable model is not yet a reliable worker.
- context budgeting
- staged missions
- tool hygiene
- static / unit / integration verification
- failure classification
- bounded retries
- checkpoints
- self-tests
- profile-agnostic behavior
- vendor independence
A capable model is not yet a reliable worker.
github.com/MajidAsghariTabrizi/universal-engineering-agent →
LONG-RUNNING INTELLIGENCE
From single prompts to missions that survive their own failures.
- single prompts → long-running missions
- tool execution with bounded authority
- provider abstraction → model independence
- checkpoints and context management
- failure recovery and reusable agent roles
Architecture level only — no private code, credentials or operational data.
WHAT IF THE SYSTEM DIDN’T START FROM ZERO EVERY TIME?
A persistent learning layer — governed, not magical.
The same loop, governed: every write carries provenance; insufficient evidence abstains.
- lexical retrieval
- semantic retrieval
- graph retrieval
- provenance
- controlled learning writes
- episodic / semantic / procedural separation
- reflection candidates
- procedure → skill projection
- shadow evaluation
- reinforcement
- decay / staleness
- rollback / rejection
- abstention when evidence is insufficient
Learning is governed. A model does not get to rewrite organizational truth because it generated a confident sentence.
THE SPLIT-BRAIN MOMENT
Two stores, one truth — until they drift.
A system cannot learn reliably if it cannot answer which memory is true.
The research loop consults prior trajectories and avoids repeating failed branches. The live trading decision loop still does not directly consume Brain memory. Two loops, honestly separated — not a self-learning trader.
HYPOTHESISA persistent learning layer makes the system smarter without becoming untrustworthy.
WHAT WE BUILTEpisodic / semantic / procedural memory with provenance, controlled writes, reflection, procedure → skill projection, shadow evaluation, decay, rollback, abstention.
WHAT REALITY SAIDA local memory snapshot and the authoritative runtime memory diverged — two stores drifting toward different truths. Resolved by explicit source-of-truth governance.
WHAT CHANGED NEXTA system cannot learn reliably if it cannot answer which memory is true.
THE CAREER BECOMES THE ARCHITECTURE
Zoom out.
Quantiviq is not a random new direction. It is the convergence of everything that came before it.
I BUILD SYSTEMS THAT LEARN FROM THEIR OWN WORK.
Technology taught me how systems run.
Product taught me what should run.
Experiments taught me that a result can be technically correct and still economically useless.
Failure taught me to separate what I built from what I could actually prove.
Organizations taught me where decisions break.
Markets taught me to measure uncertainty.
AI taught me that intelligence can become infrastructure.
Now I am exploring what happens when organizations themselves can learn.
I keep building systems that force me to learn what I did not yet understand.
VERDICTS ON THIS PAGE: PROVED · PARTIALLY PROVED · ABSTAINED · REJECTED · FAILED · UNPROVEN · PROMOTED — an honest ledger, not a highlight reel.