Nigel Misquitta@nigelnm171082450
3
Events attended
2
Submissions made
India
2+ years of experience
About me
A passionate Machine Learning & Data Science enthusiast with a strong academic background in Computer Engineering (Honors in AI & ML) from Fr. Conceicao Rodrigues College of Engineering. As a technology enthusiast, I am constantly driven to explore the ever-evolving field of Machine Learning and Data Science. My fascination lies in uncovering insights from complex data sets, applying predictive models, and creating intelligent solutions that can make a meaningful impact. I am eager to contribute my skills and knowledge to innovative organizations and collaborate with talented individuals who share my passion for leveraging data-driven approaches to solve real-world problems.
🤝 Top Collaborators
.png&w=640&q=75)
.png&w=640&q=75)
🤓 Latest Submissions
Option Alpha Agent
Agent We are building an autonomous options trading agent that runs entirely on Alpaca's paper trading environment, using Alpaca's Trading API, official MCP server, and CLI as its execution backbone. The agent identifies option-selling and directional opportunities using a two-factor decision model: implied volatility conditions (is options premium currently rich or cheap) and underlying price trend (is the stock trending or range-bound). Rather than committing to a single static strategy, the agent switches between defined-risk options structures depending on the market regime it detects, while operating inside strict, non-negotiable risk limits on every trade it places.
4 Sep 2026
Overfitters_General_Purpose_AI_Agent
Our Track 1 submission is a general-purpose AI agent built to handle eight distinct capability categories — factual knowledge, mathematical reasoning, sentiment classification, summarisation, named entity recognition, code debugging, logical reasoning, and code generation — through a single lightweight orchestrator rather than a heavyweight multi-agent framework. The pipeline works in four stages. First, a rule-based classifier assigns each incoming task to its category using deterministic keyword and structural heuristics, at zero extra token cost. Second, a model router maps each category to the right-sized model from the five allowed Fireworks models — cheap models handle sentiment and NER, a code-specialized model handles all debugging and generation tasks, and the largest model is reserved for genuine multi-step math and logical reasoning. Third, category-specific prompt templates keep every call lean, explicit about output format, and free of unnecessary chain-of-thought. Fourth, a validation and retry layer checks every response for malformed JSON, empty answers, and leaked reasoning before it's trusted, with a bounded retry policy and a guaranteed fallback so every task_id receives a non-empty answer even under total API failure. Every design decision was empirically tuned against real test logs rather than assumptions — including catching and fixing an on-demand model cold-start issue, reasoning leakage from a code-specialized model, wasted retries on non-retryable errors, and duplicate task_id handling. The result is an agent designed around how the competition is actually scored: pass the LLM-Judge accuracy gate first, then compete on raw token efficiency — not by cutting corners on correctness, but by never spending more tokens than a task genuinely needs.
13 Jul 2026
