Every company forgets
why it decided.
KAIROS never does.
Nine AI agents read your Slack, email, Drive, Notion, Jira and Zoom — extracting every decision into a living decision graph that proactively flags risk. Ask in plain English, get the answer with sources in seconds.
See It Answer
Ask the question everyone forgot.
Not a search box that hands you links — a straight answer with the who, when, and why, every claim traced to its source.
The Problem
Knowledge doesn't leave in documents.
It leaves in people.
The vendor nobody remembers
A contract signed in 2019 auto-renewed three times. The person who signed it left in 2022. Nobody knew why you were still paying.
Onboarding archaeology
A new engineer asks 'why React, not Vue?' The decision-maker is gone. The answer is buried in a 2022 Slack thread no one can find.
The mistake you repeat
'Has anyone tried a mobile app before?' Yes — it failed in 2021 for a reason you're about to hit again. The postmortem was never read.
How It Works
From raw chatter to cited answers
Connect
One-click OAuth into Slack, Gmail, Drive, Notion, Zoom, GitHub and Jira. No admin install, no IT ticket.
Extract
Ten agents read continuously, catching every decision-shaped moment with sources, people, and outcomes.
Graph
Every decision auto-links to related ones by topic, person, and timeframe — a living, physics-simulated web.
Ask
Query in plain English over chat or any MCP client — cited answers in seconds, or a warning before you repeat a mistake.
The Engine
Nine agents, running in parallel
Orchestrated with LangGraph — five own a source and extract decisions, four reason over the graph to route, retrieve, synthesize and answer live queries.
Extraction Agents
Slack Agent
Reads every channel & thread, flags decision moments, captures participants and outcomes.
Email Agent
Scans Gmail for approvals, sign-offs and escalations — links threads to the decisions they made.
Drive Agent
Parses docs, specs and proposals in Google Drive for the key choices written down inside them.
Notion Agent
Walks pages and databases recursively, extracting decisions logged in specs and wikis.
GitHub Agent
Reads pull requests and issues — with review comments and discussion — across your most active repos.
Meeting Agent
Transcribes Zoom recordings with Whisper, then pinpoints decisions, timestamps and who was in the room.
Reasoning Layer
Synthesis Engine
Fuses every source into one decision graph and answers your questions with citations.
Router
Classifies every query — search, live data, general chat, or ingest — before anything else runs.
Retrieval Engine
Hybrid semantic + keyword + graph-neighbor search, personalized to your profile and history.
Live Agent
Skips memory entirely for on-demand questions — "how many unread emails do I have?" — answered live.
New — Proactive, Not Just Reactive
KAIROS doesn't wait to be asked.
A structural scan of the entire decision graph, plus one focused model call per finding — never invented, always grounded in what's actually in memory.
Precedent Check
Before your team repeats a mobile-app attempt or re-signs a vendor, KAIROS checks memory first — and gives a punchy verdict, not a hedge.
“Yes — tried in 2021, failed from no mobile expertise. Don't repeat without closing that gap first.”
Pattern Detection
A structural scan of the entire decision graph — contradictory outcomes on the same topic, unreviewed vendor spend, one person signing off on everything.
“3 infra decisions since 2022 contradict each other — and the same person signed all of them.”
Risk Prediction
Every decision gets a live 0–100 risk score — stale, unowned, or high-impact — ranked so nothing important slips through again.
“Risk 82/100 — vendor contract, no review in 3 years, no owner on record.”
Decision Debt Score
14 decisions with no review in 2+ years
Pure SQL and graph aggregation — no model call, always live. One glance tells a story your CFO will actually read.
Connectors
Connect once. No passwords to hand over.
You sign in and connect your own accounts — no admin, no IT ticket. KAIROS keeps reading in the background and your decision graph stays up to date.
Why Not Just Confluence?
A wiki stores what you write.
KAIROS remembers what you decided.
Store the documents you deliberately sit down and write.
Mines Slack threads, email approvals, and meeting calls — where real choices live, unwritten.
Returns ten blue links and leaves the reasoning to you.
Who decided, when, why, what was rejected — synthesized and cited in one reply.
Waits, silent, until someone thinks to search it.
Scores decision debt, flags contradictions, and warns you before a mistake repeats.
Works With Your AI
KAIROS MCP
Connect Claude, ChatGPT, or Cursor straight to your company's memory. Before it answers you, your AI checks what KAIROS already knows. The moment it learns something new and important, it saves that back to KAIROS too — so the next question, from anyone, gets a smarter answer.
Remembers before it answers
Claude pulls relevant company memory before it answers anything.
Saves what it learns
Claude writes new decisions back into KAIROS the moment it learns them.
Searches by topic, person, or date
Structured search across the decision graph with full source citations.
Checks for precedent
Checks whether a new plan has real precedent — or if you're about to repeat a mistake.
Finds contradictions
Proactively scans the whole graph for contradictions, stale spend, and bus-factor risk.
Scores decision risk
Scores every decision 0–100 for staleness, ownership gaps, and unreviewed impact.
Answers live, not just from memory
Runs the same chat pipeline as the KAIROS UI and returns a sourced answer, right from Claude.
Syncs your sources on demand
Kicks off a fresh ingestion pass instead of waiting for the automatic 12-minute cycle.
Hardware Partner
Every model call runs on AMD Instinct.
Fireworks AI — KAIROS's primary LLM provider — is AMD's official inference partner. Every Fireworks call KAIROS makes (synthesis, extraction, embeddings, intent classification) runs on AMD Instinct accelerators, not NVIDIA.
| Spec | MI300X | MI350X | MI355XFLAGSHIP |
|---|---|---|---|
| Architecture | CDNA 3 | CDNA 4 | CDNA 4 |
| Process node | 5nm compute + 6nm I/O (chiplet) | 3nm (N3P) + 6nm I/O | 3nm (N3P) + 6nm I/O |
| Transistors | 153B | 185B | 185B |
| Compute units | 304 | 256 (8 XCDs × 32) | 256 (8 XCDs × 32) |
| Matrix cores | 1,216 | 1,024 | 1,024 |
| Peak clock | 2,100 MHz | 2,200 MHz | 2,400 MHz |
| Memory | 192GB HBM3 | 288GB HBM3E | 288GB HBM3E |
| Memory bandwidth | 5.3 TB/s | 8 TB/s | 8 TB/s |
| Peak FP16/BF16 (matrix) | 1,307 TFLOPS | ≈2.3 PFLOPS dense | 2.5 PFLOPS dense · 5.0 PFLOPS (2:4 sparse) |
| Peak FP8/INT8 (matrix) | 2,615 TFLOPS | ≈4.6 PFLOPS dense | 5.0 PFLOPS dense · 10.1 PFLOPS (2:4 sparse) |
| Native FP6 / FP4 | Not supported | MXFP6 · MXFP4 | MXFP6 · MXFP4, 10.1 PFLOPS |
| Interconnect | Infinity Fabric 3.0, ~896 GB/s | 7× Infinity Fabric links | 7× Infinity Fabric links @ 153 GB/s each |
| TDP | 750W | 1,000W (air-cooled) | 1,400W (liquid-cooled) |
Compute stack
ROCm, hipBLASLt, Composable Kernel
Fireworks' serving stack targets AMD's open ROCm runtime directly — vLLM and SGLang kernels compiled against Instinct, no CUDA translation layer in the path.
Native low-precision
MXFP6 / MXFP4, 2:4 sparsity
New in CDNA 4 (MI350X/MI355X) — structured 2:4 sparsity alone doubles matrix throughput, from 5.0 to 10.1 PFLOPS FP8 on the MI355X.
Chiplet packaging
8 XCDs, 3D-stacked
Every Instinct GPU KAIROS runs on is a multi-die package — compute dies on the leading node, I/O dies on a cheaper one, stitched together over Infinity Fabric.
Impact on KAIROS
KAIROS fans ten agents out in parallel every ingestion cycle — Slack, Gmail, Drive, Notion, Zoom, Jira, GitHub, plus Intent, Context, and Synthesis reasoning over whatever they pull in. That only stays cheap and fast because Fireworks' paid AMD Instinct capacity clears far more tokens/minute than a free-tier fallback would — the ingestion throttle in config.py (24 items/cycle) exists for the Groq safety-net path, not the AMD-backed primary one. Bigger HBM capacity (192–288GB) also means Fireworks can batch many users' concurrent requests on one accelerator without KAIROS ever seeing it queue.
Specs: AMD Instinct MI300X/MI350X/MI355X datasheets, amd.com. KAIROS also draws on the $50 Fireworks AI credit issued through the AMD AI Developer Program.
Model Partner — Google DeepMind
Five sizes. One family.
One of them, live in KAIROS.
Gemma 4 shipped April 2026 under a clean Apache 2.0 license — the first time Google DeepMind's open-weight family has dropped the old research-only terms. Every size takes text and image input natively; KAIROS's IntentAgent attempts the 26B-A4B mixture-of-experts first, on the AMD Instinct hardware above, for cheap, latency-sensitive query classification — dropping straight into the primary Fireworks chain below the moment that deployment isn't reachable, so a query is never blocked waiting on it.
| Model | Parameters | Context | Modalities | Role |
|---|---|---|---|---|
| Gemma 4 E2B | 2.3B effective | 128K | Text · Image · Audio | On-device — phones, laptops |
| Gemma 4 E4B | 4.5B effective (8B total) | 128K | Text · Image · Audio | Balanced on-device tier |
| Gemma 4 12B | 12B, encoder-free | 256K | Text · Image · Audio · Video | Direct linear projections replace vision/audio encoders |
| Gemma 4 26B-A4BATTEMPTED FIRST | 26B total / 3.8B active (MoE) | 256K | Text · Image | Attempted first by KAIROS's Intent Agent — dense-4B cost, MoE reasoning |
| Gemma 4 31B | 31B dense (32.2B) | 256K | Text · Image · Video | Server-grade — the top of the family |
Official Gemma 4 Model-Card Benchmarks
Score climbs with every size, across every discipline
Hover any point on the chart for the full breakdown at that size.
Source: Gemma 4 model card — ai.google.dev/gemma/docs/core/model_card_4
Impact on KAIROS
IntentAgent runs first on every single query — before Context or Synthesis ever sees it — deciding whether you asked to search memory, pull live Gmail/Drive/Jira data, or just record a new decision. Without a dedicated fast path, that classification hop would ride the same flagship model as the answer itself, paying its full latency and cost twice per question. Routing it to Gemma 4's 26B-A4B instead — 3.8B active parameters, so it runs at dense-4B speed — is designed to cut that first hop's cost and latency without giving up instruction-following reliability, and the fallback in agents/intent_agent.py drops straight back to the primary Fireworks chain the instant that deployment isn't reachable — so classification never blocks on it either way.
Reliability
One answer. Three chances to get there.
Every query starts with a cheap routing hop, then walks the fixed text-generation order below. Each step is only ever tried after the one before it genuinely fails or finishes its narrow job, so one provider having a bad minute never surfaces as a broken answer.
Gemma 4 26B-A4B
AMD Instinct, via Fireworks
The very first hop on every query — before anything else runs, KAIROS's Intent Agent tries routing classification here for dense-4B cost at MoE reasoning quality.
Attempted first; drops straight to the chain below if unreachable.
gpt-oss-120b
AMD Instinct (MI300X/MI350X/MI355X)
Every synthesis, extraction, and live-data answer starts here — plus query classification itself whenever the Gemma fast path isn't reachable. Benchmarked against every other model this account can reach; see below.
Always tried first for every real answer, no exceptions.
llama-3.1-8b-instant
Groq LPU
If Fireworks errors or hits a rate limit mid-request, KAIROS retries on Groq automatically, inside the same call — the user never sees the failure, just a slightly slower answer.
Only after a real Fireworks error.
gemini-2.0-flash
Google TPU
Closes the loop if Fireworks AND Groq both fail. Also KAIROS's primary embeddings provider — Fireworks' own embedding model is the fallback there, reversing the text-generation order.
Only if Fireworks AND Groq both fail.
Chosen by benchmark, not guesswork
Every chat model this Fireworks account can reach, tested head-to-head
6 real production queries through the exact router prompt in agents/intent_agent.py.
| Model | Avg. completion tokens | Avg. latency | Accuracy |
|---|---|---|---|
| gpt-oss-120bPRIMARY | 132 | 1.4s | 5/6 |
| glm-5p2 | 130 | 3.1s | 5/6 |
| glm-5p1 | 141 | 3.5s | 5/6 |
| deepseek-v4-pro | 150 | 3.8s | 5/6 |
| qwen3p7-plus | 750 | 3.7s | 5/6 |
| kimi-k2p6 | 315 | 4.0s | 5/6 |
gpt-oss-120b won on every axis that matters for a router: 5.7× fewer tokens and 2.6× lower latency than the next candidate, at matching accuracy — with zero truncated responses. Some larger models spend hundreds of tokens narrating a reasoning preamble before ever emitting the routing decision; on a latency-sensitive hop that runs before every single answer, that difference compounds fast.
In one sentence
KAIROS tries Gemma on AMD Instinct first for cheap query routing, falls into Fireworks' gpt-oss-120b — also on AMD Instinct — for every real answer, and only reaches for Groq or Gemini if the AMD-backed path itself is down. AMD hardware is never the fallback here — it's the path every request takes by default.
Q&A Cheat Sheet — AMD / Gemma / Fallbacks
“Does this actually run on AMD?”
Yes — the real, permanent connection is Fireworks. KAIROS's brain calls Fireworks' API, and Fireworks runs that service on real AMD Instinct hardware as their official inference partner. Every synthesis, extraction, and live-data answer goes through AMD Instinct. Separately, development happened on an AMD dev cloud machine (verified with rocm-smi) — that's just where the code was written and tested, not part of the live product path.
“What about Gemma — is that on AMD too?”
Attempted first, not guaranteed yet — and the product says so out loud, it's not hidden. Gemma 4 26B-A4B is used for one narrow job: classifying what kind of question came in, before routing it, on AMD Instinct. Fireworks hasn't enabled AMD support for that specific model architecture yet (confirmed live: rejected on MI300X, NVIDIA-only for now). KAIROS's fallback chain already handles this — when Gemma isn't reachable, it drops straight back to the same AMD-backed Fireworks → Groq → Gemini chain used for every real answer. Zero errors shown to the user, zero code changes needed the moment Fireworks flips Gemma's AMD support on.
“Walk me through your fallbacks.”
Nothing in KAIROS ever just breaks — everything has a plan B and downgrades quietly instead of erroring:
- 1Main AI brain: Fireworks (gpt-oss-120b, AMD Instinct) → Groq → Gemini.
- 2Question-sorting step: Gemma 4 on AMD first, falls back to #1's chain if unavailable.
- 3Embeddings (text → searchable vectors): Gemini → Fireworks → local basic option.
- 4Jira: per-user OAuth is fully coded, pending an Atlassian app-registration step (external, not code) — falls back to one shared login in the meantime.
- 5Zoom: same shared-fallback pattern, scoped to the owner account only — no shared fallback for other users.
- 6Notion: one-click OAuth normally, manual API-key paste as backup if that fails.
- 7Meeting transcription: skipped gracefully (not crashed) if Whisper isn't installed on that deployment.
About Me & Why Kairos
Baljot Singh (Founder & Builder)
BCA student from Punjab, India, working in agentic AI and AI automation — designing systems and directing AI tools to build them.
KAIROS exists because of a pattern seen across every project: the reasoning behind a decision disappears the moment the conversation ends. A vendor contract renews for years because no one remembers why it was signed. A new hire re-learns lessons the team already paid for. The context was never lost — it was just never captured anywhere queryable.
KAIROS solves this directly. Ten AI agents run in parallel across Slack, Gmail, Drive, Notion, Zoom, Jira, and GitHub — extracting decisions, not documents: who decided, why, what alternatives were considered, and where the reasoning lives. Ask “why did we decide X,” and KAIROS walks the decision graph and answers in under 4 seconds, sourced back to the original thread.
It connects directly to Claude, ChatGPT, or Cursor through MCP — so any AI already in use checks memory before answering, and writes new decisions back the moment they're made.
Built solo, end to end: architecture, agent pipeline, MCP server, and product, for the AMD Developer Hackathon ACT II.