Nishanth S

Stack

Python, PySide6, SQL, MVC

Type

Onsite

Stack

Python, PyTorch, Transformers, Vision-Language Models, LoRA, 4-bit Quantization, Gradient Checkpointing, CUDA

Type

Hybrid

AI-powered finance platform: a FastAPI + XGBoost prediction agent with purged walk-forward CV and conformal uncertainty bands, behind a full digital-wallet frontend.

• The model does not beat an always-up baseline on raw daily direction - 0.5264 against 0.5329 - and a cheap ARIMA beats them both. That is the finding, not a bug: daily bars are near-efficient, and every technique below exists to stop the pipeline from faking a better number.
• Refused to leak the future into training. Purged and embargoed walk-forward cross-validation, triple-barrier labeling, fractional differentiation tuned per asset by ADF test - and the leakage audit ships in the repo rather than in a claim: zero overlapping labels across 118,278 evaluation events.
• Made the model's stated confidence mean something. Expected calibration error falls from 0.0298 to effectively zero, and forecasts come out as GARCH-shaped conformal bands instead of point estimates: the 80% band realizes 80.0% coverage, the 90% band 90.0%.
• Put the edge in selectivity instead of accuracy, since the accuracy was never going to arrive. A meta-labeler decides whether to act at all, lifting precision on taken trades from 0.534 to 0.559; cross-sectional construction with vol targeting returns a net-of-cost Sharpe of 0.84 against an equal-weight benchmark's 0.75, and regime gating holds maximum drawdown to -30.8% where buy-and-hold takes -77.5%.
• Let the system overrule itself, in both directions. The LLM verifier may veto or downgrade a call but never raise it, a greedy answer is clamped to the model's own ceiling, and an outage leaves confidence untouched; a regime mixture-of-experts that lost in cross-validation ships disabled, and the stacked ensemble serves only while it beats the best single base learner.
• Ran it as a product, not a notebook. Thirteen market, macro and fundamentals providers behind optional keys, portfolio and paper trading over a Python API, deployed on Cloudflare Pages and Render with hosted inference, resolved outcomes feeding back into live calibration, a rate-limited retrain when it drifts, and 67 tests plus standing audits on two Python versions every push.

A SQL query engine built from scratch in C#, featuring a cost-based planner, a B+ tree index and a lazy Volcano-style execution model. Performance is validated via BenchmarkDotNet.

• Implemented the B+ tree index - node splitting, linked leaves, range scans. The hash index beside it is capability-aware, so the planner can never route a range predicate to an index that cannot serve one.
• Built a cost-based planner that compares SeqScan and IndexScan cost from table statistics and declines its own index once a predicate matches most of the table: 880x faster at selectivity 0.001, marginally slower at 1.0.
• Executed queries through a lazy, pull-based Volcano operator tree, so LIMIT 10 over 190k rows never materializes the full scan - roughly 8,300x faster than the eager path, and allocating by the same margin less.
• Verified correctness by differential testing against SQLite as an oracle: 500 randomly generated queries plus targeted cases for three-valued NULL logic, run on every invocation against the real dataset rather than a fixture.
• Parallelized GROUP BY aggregation across partitions with zero shared mutable state, measured across 1 to 8 threads for a 6.1x speedup.

Natural-language analytics assistant for FMCG sales data - a 5-agent pipeline that turns plain-English questions into governed DuckDB SQL and structured business answers.

• Enforced governance in the validator rather than trusting the model: generated SQL is parsed to an AST with sqlglot, checked against table and column allow-lists, stripped of DDL and DML, and LIMIT-bounded before it ever reaches the database.
• Grounded generation in a semantic layer of table and column descriptions and three certified business-metric formulas, then checked that a metric named in the question was genuinely computed through its registered formula, with a bounded repair loop feeding validator errors back to the generator.
• Gave every LLM-calling stage a deterministic fallback - a template bank for SQL, keyword overlap for retrieval, rule-based synthesis - so the pipeline still answers with no API key configured, and returned each answer with its per-agent trace and SQL so any result can be reproduced.
• Locked the behaviour down with 51 tests, including a committed 11-query gold set the orchestrator runs on every invocation, asserting both shape and values.

AI-native video editing framework built on MCP - Claude orchestrates 19 specialized tools into full editing pipelines from a single natural-language prompt.

• Replaced the manual timeline loop with a sentence. One prompt - upscale the intro, cut the character out, put them on a gradient, reveal the text on the beat drop, grade it teal-orange, export for TikTok - is decomposed by the model into the right tool calls in the right order and executed without anyone opening a timeline.
• Made that possible by shipping each capability as its own standalone MCP server: 19 servers exposing 56 tools, none of which knows the others exist. Composition is the model's job, so a new capability is a new server rather than a change to the framework, and the assistant connects to all of them at once and mixes them freely.
• Covered every stage from raw footage to finished cut - scene-cut detection and frame extraction, Real-ESRGAN upscaling, ONNX ISNet foreground segmentation, stabilization, chroma key, speed ramps, BPM and downbeat detection, compositing, text overlay, LUT grading and export presets - each one automating a task that is otherwise manual clicking.
• Drove DaVinci Resolve programmatically for timeline assembly, so the automated edit lands in a real editor a human can keep working in rather than in a bespoke renderer.

3D scrollytelling landing page for a fictional construction studio - three.js, GSAP and a React photo gallery.

• Baked the hero geometry offline - a 116 MB source model reduced to a few hundred KB of phase-tagged wireframe edges, so the house draws itself in build order.
• Built the page on one small token set - near-black ground, off-white type, a single accent, one easing curve - with three typefaces each scoped to one job.
• Lenis drives GSAP ScrollTrigger and scroll velocity feeds a global distortion shader, with all five sections sharing one WebGL stage and one draw loop.
• Reduced motion is tiered rather than switched off, and React loads only for the two photo components that need it.