>work

tl;dr

highlights:

~1% of network emissions, backed by yzi labs, dojo interface coder 7b, sft and dpo datasets, signature verification for multisig setups, kami

status:

live for 1 year 2 months, then archived

year:

2024-2025

role:

architect, product owner, sole developer → tech lead

stack:

python, go, typescript, nextjs, react, litellm, instructor, langfuse, openrouter, huggingface, playwright, typescript-lsp, postgres, milvus, redis, docker, aws, pm2, fastapi, gin, cohere.ai, siws

challenges:

game theory, incentive mechanism design, anti-gaming mechanisms in an adversarial open-source environment, embracing new ai tooling when it hasn't matured, designing cicd flows to prevent vibe coded slop

lessons:

build the product first then decentralize, narrow scope beats broad ambition, without evals every architecture decision is a judgment call, assume everything will be reverse-engineered in open-source adversarial environments, technical decisions need to be made by the people who understand the engineering

links:

synthetic api, dojo v1, worker api, messaging, kami, docs

tl;dr

highlights:

<1s time to first token (TTFT), 2000+ insights generated from 40+ bittensor youtube channels, secured USD$3m funding, inference on bittensor corcel's decentralized API with fallbacks via openrouter/litellm

status:

deployed for 1 year 6 months, then archived

year:

2023-2025

role:

architect, backend & ai engineer

stack:

python, typescript, nextjs, react, litellm, instructor, langfuse, openrouter, huggingface, postgres, milvus, redis, docker, aws, pm2, fastapi, cron, scheduling, cohere.ai, together.ai

challenges:

ai observability, writing evals, measuring system performance of reranking and query classification/decomposition/expansion, chunking experiments for sentence window/small-to-big retrieval, scraping youtube with heavy anti-botting mechanisms

lessons:

it's easier to write your own ai components than to use langchain, evals & observability are essential before measuring effects of new components in your system, evals evals EVALS

links:

release announcement, fundraise announcement

tl;dr

highlights:

deployed on Raydium, Minted, Solend, Mane City projects totalling 200k+ members, custom knowledge base management, offloaded community managers workload by 80% by answering FAQs, allowed admins to update knowledge base natively on discord

status:

archived

year:

2023-2024

role:

backend & ai engineer

stack:

python, typescript, nextjs, react, chroma, langchain, litellm, instructor, postgres, redis, kafka, docker, aws, pm2, fastapi

challenges:

balancing speed & accuracy with gpt-3.5's 15 tokens per second (TPS) bottleneck, scaling infrastructure & reliability for users, gitbook & other documentation sources may be stale

lessons:

enforce microservice architecture design to reduce code complexity, vector database requires careful consideration of embedding model, indexing, and retrieval strategies

links:

website snapshot

tl;dr

highlights:

generated signals via sentiment analysis of trader chatroom messages, performed named entity recognition against coingecko listed tokens

status:

beta test

year:

2023-2024

role:

sole developer, backend & ai engineer

stack:

python, typescript, nextjs, react, chroma, langchain, postgres, redis, kafka, docker, aws, pm2, fastapi

challenges:

named entity recognition of token names & symbols and twitter handles, context engineering of messages to handle noisy chatroom messages, reverse engineering twitter api

links:

twitter

tl;dr

highlights:

rewrote existing aggregator from python to go for 1000x speedup, aggregated order book data from Binance, MEXC, Bybit, etc.

status:

internal use

year:

2023-2023

role:

sole developer, backend engineer

stack:

go, rueidis, sonic, websockets, postgres, gin, redis, docker

challenges:

concurrency of handling 3000+ websocket connections per token across different exchanges, getting latency as low as possible to prevent stale data

lessons:

red black trees are decent for maintaining an order book, overhead from json serialization & deserialization is surprisingly high, kimchi premium exists iykyk

tl;dr

highlights:

built REST APIs for a dashboard used by 120 internal users to analyze growth channels and automate business reporting, achieved 95% faster end-to-end response times, integrated third-party CRM software to automate affiliate deal pipelines to achieve to 500% YoY affiliate growth

status:

internal use

year:

2022-2023

role:

backend engineer, scrum master

stack:

java, nodejs, mybatis, sql, git, hubspot

challenges:

translating growth and affiliate workflows into maintainable backend APIs, integrating third-party CRM data with internal systems, balancing feature delivery with test coverage and microservice uptime

lessons:

clear technical documentation compounds across a team, API performance matters even for internal tools

tl;dr

highlights:

co-led a team of 5 to deliver AIDrivers Singapore office's first computer vision product, built an instance segmentation model with over 80% accuracy to detect incorrect safety equipment usage in warehouse environments

status:

prototype delivered

year:

2020-2021

role:

autonomy engineer

stack:

python, pytorch, docker, ansible, bash, computer vision, Facebook AI Research detectron2

challenges:

building a reliable detection system for real warehouse safety scenarios, coordinating model development & data labelling across a small team, automating reproducible development and production environments with idempotency

lessons:

deployment and onboarding automation matter even in early prototypes, computer vision systems need to be evaluated against the operating environment rather than only offline metrics, simple containerized workflows reduce handoff friction