Michael Khosho: Capital Markets Quant Developer, Systems Architect, Software Developer, Product Owner

Back-end Python Developer @ BMO Capital Markets

Quant developer

Pricing, risk and market-data code in production, and a trading system I built and run.

BMOReconciled convertible dirty pricing against a second market-data vendor and had the incorrect values replaced.
BMORebuilt the nightly counterparty credit-risk pipeline: about 3 hours down to 40 minutes, output matched one-to-one.
BMOWrote the ETF decomposition feed: daily index units, weights, NAV and creation units from a vendor API.
MBATSSignal-driven execution with risk guardrails, backtests over years of one-minute history, nightly model scoring.

See MBATS

Current role

I own Python and SQL delivery on two capital-markets trading platforms: trade processing, back-office reporting, settlements, counterparty credit risk, pricing and regulatory feeds.

BMO Financial Group, Capital Markets

Backend Python Software Developer
February 2025 to present, Toronto

~500ktrades a day

60tracked initiatives

4audit cycles as control owner

Trade ingestion
Built an intraday pipeline loading front-office trade deltas every five minutes with batched Oracle MERGE upserts, encoding fallback, failed-file quarantine and a per-file audit trail.
Credit risk
Cut the nightly counterparty credit-risk job from about three hours to about 40 minutes (~78%) and removed manual restarts, validating output one-to-one against the legacy extract.
Event integration
Rebuilt a hard-coded monolithic transfer process as a config-driven, multi-entity package publishing trade events to Kafka, with separate extract and publish jobs so publishing can be held or replayed.
Pricing
Corrected dirty pricing on convertibles by reconciling zero-accrued positions against a second market-data vendor, and drove the platform vendor to replace the values.
Controls
Added a pre-batch completeness gate so the nightly back-office batch cannot run on a partial day, plus reconciliation reports and severity-tiered alerting before anything reaches clients.
Modernisation
Moved legacy .NET and JavaScript reporting and credit-risk processes into modular, config-driven Python; the reconciliation suite ran in parallel with the legacy batch before decommission.
Integrations
Oracle PL/SQL, SQL Server T-SQL, Apache Kafka, IBM MQ, REST and vendor market-data APIs, scheduled through Autosys.

Experience

Employment, and the systems I build and run on my own.

BMO Financial Group, Capital Markets
Backend Python Software Developer. February 2025 to present
Current

Python and SQL trade processing, ETL and reporting on two capital-markets platforms (swaps, loans, ETFs, convertibles, prime brokerage), about 500k trades a day. Details above.
MBATS
Private trading system. February 2026 to present
Live, private

A microservices trading system I designed, built and operate: about twenty Docker services for market data, signal-driven execution with risk guardrails, nightly model scoring, a trade journal and a research UI behind one sign-in gate.
Interactive demo Video (1:13)Architecture
resumeMaster
Automated job-search pipeline. September 2026 to present
Live

Scans LinkedIn and Indeed twice a day, matches postings against my constraints, tailors a résumé and cover letter from a database where every claim cites its source, and applies through the browser. Unknown screening questions come to me by Telegram with a suggested answer. This site is generated from the same database.
Video (1:08)Architecture
ATLAS
Multi-agent business automation. August 2026 to present
In development

A Python platform for delivering AI automations to small service businesses, where a second client is configuration rather than new code. Architecture.
Revature
L2 Production Support and Software Developer. September 2024 to present, contractor: CitiBank, then BMO
- L2 production support in ServiceNow for stock-market trading tools and applications.
- Trading tools for traders in US and Canadian stocks, securities and bonds.
- RESTful APIs with Python, Flask, SQLAlchemy and Azure DB.
- CI/CD with Ansible, Jenkins and Pytest; messaging with RabbitMQ and Celery.
Horizon Digital
Software Developer and Data Analyst. July 2021 to September 2024, remote
- Designed, developed and deployed web applications with cross-functional teams: HTML, CSS, JavaScript, Angular, SOAP, .NET, SQL and Git.
- Agile delivery with modular codebases.
University of Windsor
BSc, Computer Science

Web systems development, operating systems, object-oriented programming in Java, data structures and algorithms, computer architecture.

Under the hood

How each system is put together, drawn at the level of its architecture. Strategy rules, results, client work and internal names stay private.

MBATS, a private trading system

Live, privateML gating and sizing: planned

Personal project, February 2026 to present. 348 of 372 commits and 106 merged pull requests are mine; the rest is the open-source base it started from. The strategy was specified by a collaborator. Access is limited to its two operators; the public demo runs on synthetic data.

Interactive demo

Walkthrough of the public demo, 1 min 13 s, no sound. Every figure on screen is synthetic demo data.

Architecture only. Strategy rules and results are private.

Two planes: research changes freely while execution stays fixed; new behaviour ships switched off behind pre-registered gates.
A direct webhook replaced a chat-relay hop, removing about one to three seconds and a class of silent failures.
Model scoring fails open and execution fails closed: a missing score never blocks or alters an exit.
No ML stack inside the execution container; models score nightly and execution only attaches the result.
543 tests, including a regression suite numbered by incident, and a 75-incident troubleshooting log.

Python, FastAPI, Pydantic, SQLAlchemy, PostgreSQL, Airflow, MLflow, XGBoost, Optuna, PyTorch, Vue 3, React with TypeScript, Docker Compose, nginx, Firebase Auth

resumeMaster, an automated job-search pipeline

Live

Personal project, September 2026 to present. I designed it and directed the build; much of the code was written with Claude Code. 41 commits, 10 pull requests, 97 tests.

Architecture and a tour of the control panel, 1 min 8 s, no sound.

The loader refuses any record without a source, so nothing it generates can claim more than the records show.

One database of record with fail-closed provenance: every highlight and skill must cite the document it came from.
Applications stop at the review step unless auto-submit is switched on; a kill switch pauses the agent.
Questions the forms ask that it cannot answer from records go to me with a suggested answer; one reply answers the same question worded differently.
Everything is local and inspectable: SQLite, plain files, one command-line entry point.

Python 3.12, SQLite, FastAPI, Playwright, ReportLab, Telegram Bot API

Intraday trade ingestion

In production at BMO

Software developer, October 2025 to September 2026.

Front-office trade deltas arrive every five minutes. Each file is tracked in a payload table, merged into Oracle in batches of 250, and moved to quarantine if it fails, with a fallback across encodings. A completeness gate holds the nightly back-office batch until the day is whole.

Outcome: five-minute loads with per-file auditability, duplicates eliminated, and encoding failures no longer stall the queue.

Python, Oracle, Autosys

ATLAS, multi-agent business automation

In development

Personal project, August 2026 to present. I designed the system, wrote the specification and the core, and directed AI-assisted implementation of several subsystems. Over 700 tests. Nothing is deployed for real users yet.

Taking on a second client is configuration rather than new code. Every stage is stateless, auditable and gated by deterministic checks.

Architecture only. Client work and business details are private.

All state and memory live in two places, so every other agent is stateless, testable alone and safe to re-run.
The orchestration framework stops at the orchestrator; no shared skill depends on it.
No language model inside a gate, a solver or the runtime. Those cores are deterministic.
Provenance is enforced by the vault schema: every claim needs a source or a derivation.
CI never calls a model: tests replay recorded responses, and an unbound provider is an error, never a fallback.

Python 3.12, Pydantic, LangGraph, SQLite, OR-Tools CP-SAT, pytest, ruff, mypy strict, GitHub Actions