Case studies

Problem

Teams that cannot send data to a cloud API still want agents, retrieval, voice and document tools. The stack has to run offline on one GPU box and stay up while several models compete for memory.

The stack

• Gateway. FastAPI with an OpenAI-compatible /v1/chat/completions and /v1/embeddings, API-key RBAC, and routing across vLLM, ONNX Runtime, ComfyUI and llama-swap backends.
• GPU memory manager. Estimates VRAM before loading, evicts the least recently used model, queues requests, and caps concurrency.
• Code sandbox. Agent-written Python, JS, Bash and R run in Docker with gVisor, no network and CPU and memory limits.
• Client. Tool-calling agent orchestrator, React dashboard, CLI with server pairing and a self-verification suite, offline voice, and local DOCX, PDF, PPTX and XLSX generation.

Failure modes designed for

An out-of-memory crash when a second model loads, so VRAM is estimated first. A burst of requests, so a queue and a concurrency cap sit in front of the GPU. Untrusted generated code, so it never touches the network. Restricted hosts that cannot run gVisor, so the sandbox has a degraded mode instead of failing to start.

Code

Server and client. Team project for SIH 2026.

Question

If a model is told that one answer suits a stakeholder, does its numerical estimate move? A correct estimate should not depend on who needs which number.

Method

Three Fermi problems (illegal logging in a national park, security remediation hours, unpaid overtime among gig workers), each asked neutrally and with a framing that favours a higher or lower number. Run on ChatGPT, Gemini, Qwen and Claude.

Finding

Estimates shifted by roughly 2x to 5x depending on framing across the models tested. One model’s estimate moved from about 8,000 to about 20,000 across framings.

Limits, stated plainly

Small sample, so ordinary run-to-run variance is not separated from the framing effect. The next step is 5 to 10 runs per condition per model and a neutral-run spread to compare against. I treat this as a lead worth testing, not a result.

Problem

People fill the same details into application forms again and again, on sites whose structure nobody controls.

Design

The user gives a profile once. An agent finds the target form, infers the required fields and constraints from the page, then generates browser automation to fill it. CivicFlow adds Claude Vision for pages that resist parsing and a Telegram bot as the way data arrives. FAR AWAY 2026 version: code.

Takeaway

Splitting field inference from execution makes failures inspectable: when a fill goes wrong, you can see whether the model misread the form or the script misclicked.

Problem

Few samples, heavy class imbalance, and a distribution shift between the localizer and imagination recordings.

Approach

Elastic Net decoder, manual SMOTE oversampling, robust scaling, domain adaptation from localizer to imagine sessions, features from a 200 to 800 ms post-stimulus window, and an ensemble with bootstrap voting to steady predictions. Best public score 0.122.

Approach

Instead of one conventional model, I ran several LLM calls concurrently and combined their outputs. It reached 99.999 accuracy on the validation set.

Ro-SVD

A 7-configuration sweep over 1,602 samples on all 16 LongBench subsets tracking ROUGE-L, perplexity, latency and GPU memory. The Python-loop SVD cost 29.5% latency, which motivated a batched, randomised path.

Sentinel

Five attack categories, with a modality-gap extension comparing text, image and video delivery of the same attack, judged with McNemar and Wilcoxon tests.

ZENSO

An ablation showed that 16 sampled frames stay above 0.99 Pearson r against full-frame input at a 7.5x inference speedup.