Backend systems should be like Formula 1 cars: relentlessly optimized,
precisely engineered, and safe enough to push to the limit.
Race history
Three stints · 2023 to present
-
S1 · Current
Oct 2024 → now
Backend Engineer
QubitGlobal / Quantitative Trading and Analytics · Dubai, UAE (remote)
- Replaced fixed-spread quoting with an async grid market maker in Python on the GLFT model, extending Avellaneda-Stoikov: half-spreads skew with live inventory to cut adverse selection and pull position back to flat. Live on 30+ pairs across 10+ venues.
- Made spreads adaptive by estimating volatility, order-flow intensity and book depth online at 1,000 quote updates per second, removing roughly 10 manual re-tunes per week.
- Cut new-venue onboarding from 3 days to 1 by moving the FastAPI gateway to hexagonal architecture, where each adapter (Bybit, BitMart, MEXC) owns its REST and WebSocket clients, normalizer and error mapping, including binary protobuf decoding.
- Guaranteed exactly-once accounting across 100+ restarts with zero drift, on Redis-backed state and TimescaleDB persistence, plus fail-stop startup reconciling orders and balances against exchange truth in 20 seconds.
- Ran every client account from one process by multiplexing public and private authenticated streams across 13 venues, staying alive through broker outages behind a per-exchange circuit breaker that sheds rather than buffers.
- Cut bot rollout from 15 to 30 minutes down to 30 seconds with a Go control plane provisioning workloads via the Kubernetes API, a per-bot Go sidecar supervisor, a NATS control bus using wildcard subjects for per-bot addressing, RBAC, and a reconciler loop converging cluster state to desired state. 50+ bots live.
- Made environments reproducible with Terraform on DigitalOcean, including the single static egress IP that exchange API key allowlisting depends on, so every client account routes through one whitelisted origin.
- Delivered the predecessor TypeScript engine behind one connector interface, with Redis-backed state and SSE streaming.
-
S2
Jan 2024 → Sep 2024
Backend Developer
TokenPilot / Digital Asset Liquidity and Market Making · Singapore (remote)
- Integrated order management, market data and reconciliation across 10+ centralized exchanges including Binance, Bybit, MEXC, Gate.io, KuCoin and XT at 100,000 orders per day, handling rate limits, request signing and depth parsing.
- Handled live-flow failure modes: stale price detection, partial fill reconciliation, duplicate-request guards and local-versus-venue state drift.
- Built scheduled Python profit-and-loss, exposure and fill-rate reporting for 20 client accounts, plus a Telegram bot for position alerts.
-
S3
Jun 2023 → Nov 2023
Backend Intern
LancemeUp / Kathmandu
- Built REST and GraphQL APIs with NestJS, PostgreSQL and TypeORM plus a WebSocket layer for live updates, deployed with Docker, nginx and PM2.
Fastest lap
Spreads that retune themselves on a thousand quote updates a second
The engine estimates volatility, order-flow intensity and book depth online, then skews
its half-spreads against live inventory. It replaced fixed spreads, and it replaced the
engineer who used to retune them ten times a week.
Lap deltas
Before and after, measured in production
-
Time to connect a new exchange
3x faster
Before
3 days
After
1 day
-
Time to deploy a bot to the cluster
30x faster
Before
15 to 30 min
After
30 sec
-
Manual spread re-tuning by an engineer
Eliminated
Before
~10 per week
After
None
- Quote updates
- 1,000/s
- Orders handled
- 100k/day
- Pairs quoted
- 30+
- Venues connected
- 13
- Bots in cluster
- 50+
- Restarts, no drift
- 100+
- Startup reconcile
- 20s
Garage
Built outside work hours
GharBhada
Django · Next.js · PostgreSQL · PostGIS · Celery · nationwide rental platform
ghar-bhada.com →
- Built a 19-source scraping pipeline behind a plugin registry, normalizing heterogeneous HTML from 3,000+ listings into one canonical schema with URL and image dedup.
- Designed geospatial search on H3 hex indexing at resolutions 7 to 9 with PostGIS map clustering, powering map-bounds queries and locality pages generated from cell aggregates.
- Used weighted PostgreSQL full-text search instead of Elasticsearch, and self-hosted geocoding behind a fallback chain to commercial providers.
- Shipped a saved-search alert engine with push, Telegram and Discord delivery, idempotent and deduplicated, running on Celery beat alongside scraping, cleanup and backup jobs.
- Hardened the platform with Firebase and JWT auth, Turnstile bot gating, upload content-type validation, per-endpoint rate limiting, and automated Postgres backups to Cloudflare R2.
- Built a byte-serial hardware decoder for NASDAQ TotalView-ITCH 5.0, covering 9 message types and 98% of session traffic.
- Designed the framer with dual length-source cross-checking, plus a direct-mapped order book with collision detection and price-ladder occupancy tracking.
- Verified against Python reference models on real session captures: 264M+ messages byte-identical, zero frame errors, zero length mismatches.
- Splits large files into AES-256-GCM encrypted, SHA-256-verified chunks across multiple cloud providers, with round-robin allocation and manifest-based reassembly.
Debrief
Writing, newest first
Parsing NASDAQ ITCH on an FPGA
7 August 2026 · Medium
Read →
How I used Zero-Copy To Achieve Blazingly Fast File Transfers
6 August 2025 · Medium
Read →
Vertical vs Horizontal Scaling
17 December 2023 · Medium
Read →
Setup sheet
Highlighted entries are the daily drivers
- Languages
-
PythonTypeScriptGoSQLSystemVerilog
- Backend
-
asyncioFastAPIDjangoNode.jsNestJSNext.jsReactCeleryRabbitMQNATSRedisWebSocketsprotobufhexagonal architecture
- Data
-
PostgreSQLTimescaleDBPostGISH3 geospatial indexingfull-text searchMongoDB
- Infrastructure
-
KubernetesDockerTerraformDigitalOceanAWSLinuxnginxCI/CD
- Domain
-
Market makingGLFT and Avellaneda-Stoikov quotingInventory riskExchange connectivityOrder lifecycle and reconciliation