Back

maxim

maxim is an ai observability and evaluation platform — logging, tracing, evals, test runs, prompt management, and simulations for teams building on LLMs. i have worked across the whole stack — the main platform (next.js monorepo on Nx), and both Python and JS/TS SDKs.

internal bifrost deployment

maintains internal bifrost instance that powers maxim's ai layer. it allows us to do things like:

BYOK (bring your own key), provide provider-specific configs (like vertex's regions, bedrock arn, etc.), multiple deployments for Ollama and vLLM.
provider and model discovery — maxim supports 1000+ models, across 20+ providers.
inference, voice simulations, and streaming all go through bifrost.

human & retro evals

external human raters on log repositories, a variable catalog for evaluator suggestions, and the human-eval sheet/table UI (annotation forms, comparison views, xlsx export for human evaluations in test runs).

logging pipeline

consolidated the logging surface into a single logging API, added a LogLine class in the JS SDK for manually pushing logs, and fixed a nasty bug where multiple Maxim SDK instances on the same API key stepped on each other.

added the LogLine apis to collect log lines and push them to the logging apis. this gives client ownership of the log export process.

kept both SDKs in lockstep — prompt id / prompt version passthrough, variable mapping, withLogger/with_logger for logging on prompt runs, streaming fixes for the agno integration (python), and making sure a broken log-repository connection doesn't block logger creation.

prompts, evaluator, simulation apis

built out a big chunk of the public API surface — prompt tools, prompt versions v2, prompt deployment by version number, prompt partials, and model / evaluator management apis. Also shipped variable mapping for SDK test runs and prompt-version fallback logic when creating new versions.

ui work

session table improvements (tool-call display, saved search, tooltips for cost / tokens / prompt versions)
drag-to-reorder and filtering in trace and session tables
dataset split search and migration fixes
dashboard and email-template fixes and a long tail of UI bugfixes (dropdown state, title race conditions, drag-and-drop, rich text handling).

bifrost

i joined early, before stable v1 and i've stayed on the core team since, shipping across pretty much every layer — provider integrations, the compat plugin, observability and helm configs. this is what I've mostly worked on:

new providers

added elevenLabs (speech + transcription), groq STT/TTS, and most recently deepseek as a first-class provider.

compat plugin

this is the layer that makes "any client, any provider" actually work. it roughly does this:

convert one request to another and back to the original (for example, if a model does not support chat completions but supports responses, it will convert the request from chat to responses, hit the model and convert the response back to chat completions format)
automatically drops or converts unsupported parameters for each model based on bifrost's datasheet. compatibility fixes now don't need a new version release, it only needs a datasheet update (and older clients stay compatible longer).
a long tail of per-provider quirks — anthropic rejecting trailing-assistant messages, cache_control handling for bedrock, conversion of role developer, flattening of namespaced tools for non-openai providers.

multi-deployment support

reworked Ollama and SGLang from a single provider-level base URL to per-key URLs, so you can load-balance across multiple local instances serving different models.

multiple OTEL collector profiles

Same idea later for OTEL: went from one collector to a profiles array so traces can fan out to multiple destinations (Jaeger + Datadog + whatever else) at once.

pricing correctness

fixed streaming cost/usage attribution more than once (vLLM's --skip-pipeline output, image generation/edit streaming, virtual-key-scoped pricing overrides not propagating through the streaming accumulator context).

helm charts

tightened helm schema

core prs

some of my prs:

#824 — feat: support for ElevenLabs provider (speech, streaming, transcription)
#886 — feat: time-based log filtering and log deletion
#892 — feat: enable logs UI when a log store is connected
#2039 — fix: move text→chat conversion from plugin to provider dispatch level (LiteLLM compat)

#2099 — feat: Groq speech synthesis and transcription (STT/TTS)
#2102 — fix: handle text/vtt/srt response formats for transcriptions

#2386 — feat: per-key URLs for Ollama/SGLang to support multiple deployments

#2784 — fix: token usage attribution for vLLM --skip-pipeline
#2858 — feat: Codex compatibility — namespace tool type + web-search tool dropping
#2913 — feat: namespace tool type support, flattened for non-OpenAI providers
#2938 / #2965 — fix: drop empty thinking blocks that made Claude Code choke on Anthropic
#3074 — fix: fall back to provider-level network_config if per-key URL isn't set
#3154 — fix: drop cachePoint/cache_control for models/providers that don't support prompt caching
#3200 — fix: developer→system/user role handling across Anthropic, Bedrock, Gemini
#3202 — fix: compat plugin booleans were silently defaulting to false; now default true
#3203 — fix: strip trailing assistant messages for Anthropic, sanitize Mistral reasoning_effort
#3221 — fix: reasoningConfig field naming + message sanitization for non-Anthropic Bedrock models
#3421 — fix: virtual-key/provider-key pricing overrides weren't propagating through streaming context
#4630 — feat: drop reasoning when tools are present but the model doesn't support both together
#4657 — fix: strip assistant reasoning_content for Cerebras (unsupported field)
#4802 — fix: cost/usage attribution for image generation & image-edit streaming
#4852 — feat: add DeepSeek as a first-class provider
#4861 — fix: disable DeepSeek thinking mode when tool_choice: required is set
#4956 — feat: Anthropic Files API — content_type forwarding + file_id references