🤓 Latest Submissions
Nonius — The Interview Companion
Nonius is a voice interview companion for journalists: the next question, with the source attached. It helps a reporter notice useful follow-ups while retaining control of the conversation and editorial judgment. AssemblyAI real-time transcription turns microphone or WAV audio into a live conversation. Detected voices appear as Person A, B and C; names and roles apply throughout the interview. Consecutive speech stays in one card. Changed speaker attributions are proposed for review rather than silently replacing the reporter's record. AssemblyAI LLM Gateway considers the reporting brief, prepared questions, their coverage, recent speaker passages and earlier suggestions. It proposes a question category with literal supporting evidence. Application code validates the excerpts and supplies a direct question. The reporter can request follow-ups by button or voice, or enable automatic suggestions. Nonius speaks only when invited. The evidence notebook connects saved excerpts to original audio and surrounding context. Reporter notes remain separate from source words. A downloadable ZIP includes the transcript, notes, available audio and a SHA-256 manifest. Exact text and file checks establish consistency, not truth or recognition accuracy. Primary research on journalistic dialogue and AI-assisted interviewing informs the design. Reproducible software checks cover quotation integrity, speaker revisions, context, playback, exports and usage controls. Historical model measurements, failures and their narrow scope are documented in the repository. Journalist usefulness and distraction remain unvalidated; a counterbalanced practitioner pilot is planned. Independent reporters and small editorial teams are the proposed users. A per-reporter subscription is a business hypothesis. A labeled fictional sample works without API credits. Original project code is MIT licensed, with third-party licence notices preserved.
30 Sep 2026
Talos: Learned Dinner-Table Robotics
Talos turns supported spoken or typed instructions into dinner-table manipulation by two SO-101 arms in MuJoCo. Six learned skills place a bottle, plate and mug, open a passive drawer, and retrieve the fork and spoon. Speechmatics transcribes commands; a bounded language interpreter and classical RGB checks select actions. Neural motor trajectories use camera features and joint feedback. A separate simulator-state monitor stops or scores execution. The original learned dinner system passed 8/10 frozen sequences. Table-supported bottle relays passed 5/5 frozen trials in each direction; a composed relay-plus-dinner workflow passed 8/10. All failures remain. Later stereo correction during late mug placement passed 20/20 narrow workflows, as did the original baseline; frozen corrective images passed 0/20. This establishes bounded visual feedback, not general workspace coverage. A human “Set the table” voice command completed all six skills. The hosted live-mug CPU demo also passed. Six-skill baseline execution is recorded on a legacy Intel laptop with Intel CPU/iGPU inference and UHD rendering. Core Ultra Series 2/3 eligibility remains unresolved; the latest live-mug profile needs separate Intel verification. Broader work includes seven loose items, full tabletop yaw and stable alternative orientations. No complete jointly randomized seven-item scene has passed. Calibrated RGB-D and an adapted CLIPort runtime work on the RTX 4070, but no shared Talos spatial policy has been trained. Arbitrary reachable positions, general visual recovery, unrestricted language, pouring and airborne handoffs remain unfinished. Original Talos contributions are MIT licensed; third-party components retain their own licenses. Source, compact training inputs, model provenance, protocols and successful/failed physical outcomes are preserved.
16 Sep 2026
Delphi: Conformal Risk Control Condor
Delphi is an autonomous options agent on Alpaca paper trading. Each cycle it prices a 0DTE SPY iron condor and asks: does the market pay more for the band than the band is certified to pay out? The short strikes sit at or beyond a radius certified by conformal risk control on 620 historical sessions (buyer's expected payout at most 10% of the wing). The trade goes only if credit/wing at the expected fill is at least 0.15: the certified level plus a 0.05 cost margin. Around it sit 31 pre-registered hard gates, from sizing to an independent reconciliation through the Alpaca CLI. AI: open-weight models on Featherless (DeepSeek-V3.2, Qwen3-30B) return categories only: a three-call regime vote (disagreement is abstention), a critic verdict, the journal prose. Every number that reaches an order comes from code; every decision lands in an append-only audit log with the git and config hashes that produced it. A formal core of 13 theorems is proved in Lean 4. Alpaca: Trading API via alpaca-py (chains, multi-leg orders, positions, IEX bars), the CLI as a second, independent view of the account, the MCP server for operator inspection in Claude Code. Paper trading is enforced at start-up. Judged account PA31SEVJV9P9 (new, $100,000, options level 3), 2026-09-03: the market offered 0.005 to 0.0725 of the wing for the certified band against the 0.15 gate, so no order was sent. At the close SPY stood outside the certified interval; the band would have paid out 0.338 of the wing, four times the best credit offered. Zero trades, P&L $0, and the refusal was the right decision. A one-contract pilot the day before, on a separate development account, exercised the full path (+$14) and is shown labelled, not counted. We claim no Sharpe ratio and no edge; with a handful of sessions those are noise. We claim a certified bound, a risk process that behaved exactly as specified, and a ledger anyone can recompute with one script (scripts/reproduce.py). MIT.
4 Sep 2026
