photowhips
A two-way marketplace where owners list classic and exotic cars for photoshoots, film sets, and events. Past 100 vehicles, with traffic climbing.
software engineering studio
We make large language models small and precise.
bobatron is a software engineering studio specialising in llm fine-tuning. We take general-purpose models and fit them to a single job: fewer tokens, lower latency, a loss curve that flattens where it should.
train / validation lossconverged
base model70b
final loss0.31
tokens saved94%
run cost$41
services
three things, done properly- svc.01
llm fine-tuning
Distillation, supervised fine-tuning, and preference optimisation. We take a general model and fit it to one job, then measure the loss curve and the invoice until both flatten.
- sft
- dpo
- distillation
- quantisation
- svc.02
agent infrastructure
Tool routing, context budgets, and failure handling for systems that run unattended. Most agent cost is context, so most agent work is context engineering.
- tool search
- context reduction
- evaluation
- routing
- svc.03
product engineering
Next.js front ends, edge runtimes, and the databases behind them. Shipped to production and still maintainable a year later.
- next.js
- cloudflare workers
- postgres
- supabase
selected work
things we built and shippedleadscreen
betaA CRM for lead-gen shops: one drop-in script, then lead status and revenue wired back to Google Analytics and Google Ads. Beta software, use at your own risk.
kartochki
A browser language trainer pairing spaced repetition with speech checks, so you practise listening and speaking rather than only reading. Open a live deck:
MIT · underdown/flashcard
detect-bot
in progressOur playground for visitor intelligence. We squeeze as much signal out of a session as we can — the opposite of code golf, for browser fingerprinting. A release is on the way.
hermes-token-logger
open sourceA Hermes Agent plugin that logs every API call to gzip-compressed CSV: token counts, cache hit rates, estimated cost, latency. Query with /token-summary or analyse with shell one-liners.
hermes-model-pricing
open sourceLive per-million-token pricing for 4,000+ models across 100+ providers — OpenRouter, Anthropic, OpenAI, NVIDIA, Bedrock. Built into the Hermes CLI as /model and /cost.
lighthouse badge
A GitHub-style badge in the footer that runs a real Lighthouse audit through Google PageSpeed — performance, accessibility, best practices, and SEO for this site. No CLI required.
writing
all 174 posts →Hermes Agent plugins now declare and self-install Python dependencies
Plugin Python deps now auto-install with constraint resolution, conflict refusal, and survival through hermes update. 785 LOC, 17 files, 587 tests passing.
Hermes Agent Ships Plugin Catalog With 100 Pre-Vetted Extensions
65 curated MCP servers and 96 community plugins, all Nous-reviewed, with interactive install and selective tool enabling.
Hermes `tools.stub_mode` reduces non-core tool schemas to stubs
Hermes `tools.stub_mode` reduces non-core tool schemas to stubs, cutting ~2,756 tokens per call. With 25 MCP tools loaded, trajectory compression is deferred by about 15 turns.
Cutting Hermes Agent overhead with prompt-size and on-demand tools
A recent X thread broke down how to measure Hermes fixed token cost, then reduce it by disabling skills, trimming memory, and switching tool_search to auto.