.png&w=3840&q=75)
Application
RL Introductory Hackathon for envs Cartpole, Walker , Lunar Lander
Category tags:Simple Video captioner
A two-stage video captioning agent, built reliability-first. Stage one samples keyframes from each clip and sends them to a vision-language model (Kimi K2.6 on Fireworks), which returns a dense, factual description — this is what the accuracy score is made of. Stage two turns that description into the four requested styles as a single JSON object — this is what the style score is made of. Splitting them means a caption can only be as accurate as what was actually seen, and the pressure to be funny never contaminates the grounding. Everything underneath is about not failing. A complete, valid results.json is written before the first network call, then each clip is atomically replaced as it finishes, so a crash or a hang still leaves a scoreable file behind. The agent enforces its own 540-second deadline and exits cleanly rather than waiting to be killed. Each stage walks a fallback ladder — API, then local model, then template — so no single failure is fatal. In testing, a real local-model outage still produced twelve valid captions with 292 seconds to spare. The local model is the part I most wanted to build. I distilled the teacher's style behaviour into Gemma 4 with LoRA on teacher-labelled examples, merged the adapter, and quantized it to a Q4_K_M GGUF — 5.3 GB, 2.86x smaller than bf16 — served by llama-server on CPU inside the container, no GPU required. It styles a clip in about 30 seconds on two cores. It ships as the fallback rather than the primary, because I measured instead of assuming. A blind, order-swapped LLM judge (n=19, minimum detectable effect +/-0.44) found the LoRA bought style (+0.61) but cost accuracy (-0.29), and the teacher still beat the student 10-3 on swap-surviving verdicts. The same judge found Q4_K_M statistically indistinguishable from bf16 — 4-bit quantization was free. So the distilled model earns its place by surviving an outage, not by going first.
Neotron
AIVE-Artificial Intelligence Venture Engine
AIVE (Artificial Intelligence Venture Engine) is a Cognitive Discovery Operating System that transforms documents, research papers, and data into structured knowledge, evidence-backed insights, dynamic reasoning, and actionable innovation opportunities.
Quacky Wonderland
SOAP Copilot: AI Clinical Scribe on AMD
SOAP Copilot turns raw doctor-patient conversations into structured SOAP notes, ICD-10 codes, and patient-friendly summaries in seconds, using a 3-agent Llama 3.3 70B pipeline built and fine-tuned on AMD hardware.
LoneSoloWolf
SmartBandit Router
A reinforcement-learning router that learns which AI model to call for each query, cutting cost by ~93% without ever hardcoding a routing rule.
nyan
Thymus
Thymus is a lightweight hybrid token-efficient router designed to maximize accuracy while minimizing token costs in multi‑task LLM pipelines. It dynamically routes user queries across local and remote models on LLM providers.
The Disappointer
.png&w=96&q=75){kind=small}
For Innovators & Creators
Simple Video captioner
A two-stage video captioning agent, built reliability-first. Stage one samples keyframes from each clip and sends them to a vision-language model (Kimi K2.6 on Fireworks), which returns a dense, factual description — this is what the accuracy score is made of. Stage two turns that description into the four requested styles as a single JSON object — this is what the style score is made of. Splitting them means a caption can only be as accurate as what was actually seen, and the pressure to be funny never contaminates the grounding. Everything underneath is about not failing. A complete, valid results.json is written before the first network call, then each clip is atomically replaced as it finishes, so a crash or a hang still leaves a scoreable file behind. The agent enforces its own 540-second deadline and exits cleanly rather than waiting to be killed. Each stage walks a fallback ladder — API, then local model, then template — so no single failure is fatal. In testing, a real local-model outage still produced twelve valid captions with 292 seconds to spare. The local model is the part I most wanted to build. I distilled the teacher's style behaviour into Gemma 4 with LoRA on teacher-labelled examples, merged the adapter, and quantized it to a Q4_K_M GGUF — 5.3 GB, 2.86x smaller than bf16 — served by llama-server on CPU inside the container, no GPU required. It styles a clip in about 30 seconds on two cores. It ships as the fallback rather than the primary, because I measured instead of assuming. A blind, order-swapped LLM judge (n=19, minimum detectable effect +/-0.44) found the LoRA bought style (+0.61) but cost accuracy (-0.29), and the teacher still beat the student 10-3 on swap-surviving verdicts. The same judge found Q4_K_M statistically indistinguishable from bf16 — 4-bit quantization was free. So the distilled model earns its place by surviving an outage, not by going first.
Neotron
AIVE-Artificial Intelligence Venture Engine
AIVE (Artificial Intelligence Venture Engine) is a Cognitive Discovery Operating System that transforms documents, research papers, and data into structured knowledge, evidence-backed insights, dynamic reasoning, and actionable innovation opportunities.
Quacky Wonderland
SOAP Copilot: AI Clinical Scribe on AMD
SOAP Copilot turns raw doctor-patient conversations into structured SOAP notes, ICD-10 codes, and patient-friendly summaries in seconds, using a 3-agent Llama 3.3 70B pipeline built and fine-tuned on AMD hardware.
LoneSoloWolf
SmartBandit Router
A reinforcement-learning router that learns which AI model to call for each query, cutting cost by ~93% without ever hardcoding a routing rule.
nyan
Thymus
Thymus is a lightweight hybrid token-efficient router designed to maximize accuracy while minimizing token costs in multi‑task LLM pipelines. It dynamically routes user queries across local and remote models on LLM providers.
The Disappointer