AI Enterprise SolutionsFinalistGPT & Claude

Clarity - Customer Intelligence

Clarity turns customer complaints into evidence-backed investigations with a clear owner. User: support and CX managers at digital banks and e-commerce companies handling payment failures, delayed refunds and account-access problems. Problem: the same failure appears across calls in different words. Support teams handle individual conversations, while engineering needs a consolidated problem, reliable evidence and a clear investigation. That gap allows recurring issues to remain unresolved. Solution: upload a recording. Clarity automatically detects the language, transcribes speech, masks personal details, identifies issues and connects related findings. Each proposed investigation links to timestamped source evidence. Reviewers can replay the exact audio, assign responsibility and approve or reject recommendations with a recorded rationale. No manual call details are required. Measured advantage: in an eight-case synthetic transcript benchmark, Clarity matched 7 of 7 labeled issues, compared with 3 of 7 for a fixed keyword baseline. That is 2.3× as many matched issues: automated label-matching recall increased from 42.9% to 100%, with 94.5% precision. Human review handles uncertain findings and the extra prediction. Working product: the complete application is live on Azure, with fresh audio processing, local AI, persistent storage and durable workflows. Teams can start with existing recordings without a required CRM integration. Feasibility: models run on the server without external inference API fees. Quoted Azure compute cost is $0.012/hour, plus storage and public IP. Competitive distinction: Clarity connects source audio, recurring complaints, company-specific instructions and human decisions in one traceable workflow. Every approved investigation has evidence, an owner and a recorded decision. Start today: upload a call, inspect the evidence and turn a customer complaint into an actionable investigation.

Open the demo Video GitHub

Preliminary score

76.5/ 100Strongaverage of 2 judges

Value for the user19.5 / 25
A specific problem and a clear outcome from solving it.
Prototype and use of AI25.0 / 30
A working core scenario, and what AI actually contributes to the result.
Quality testing15.5 / 20
Tests, examples of failures, and comparison with the current approach.
Feasibility10.5 / 15
Data requirements, running costs and a clear next step.
Originality6.0 / 10
A meaningful difference from common solutions.

A finalist: the live jury scored it 80.0 / 100 on stage.

What the judges wrote

GPT-6 AstraPreliminary75/100
Strengths
Your documented fresh-upload run and evidence-linked review screens support a credible recording-to-investigation workflow for support and CX managers. Your local AI pipeline contributes transcription, structured findings and issue grouping, while people retain decision ownership. Your synthetic keyword comparison, failure-driven fixes and explicit distinction between prepared outputs and fresh inference provide useful evidence without claiming proven business impact.
To improve
Your next evaluation should use consented, noisy local-language recordings, independent semantic labels and a timed comparison against manual review. Your submission should reconcile conflicting figures: 10.2 versus 205.2 seconds for fresh processing, $0.012 versus $0.252 per compute hour, and isolated versus shared guest workspaces. Your current shared deployment needs its own access-control acceptance tests and a clear pre-upload disclosure that recordings are public and raw audio is not redacted. Your operating-cost evidence should include sustained throughput and total cost per processed audio hour, including storage and human review.
Claude Opus 5.5Preliminary78/100
Strengths
Your prototype is deployed end to end on Azure. A fresh recording goes through local transcription, masking, Qwen-based findings, embedding-based grouping and a human decision lifecycle, so AI is essential to the result rather than decorative. Your testing is unusually honest and layered: 131 automated tests, a deployed acceptance run with timings, and a controlled comparison against a frozen keyword baseline (7/7 vs 3/7). You also show concrete failure-driven fixes, such as per-finding evidence validation and rejecting agent-only evidence. You consistently separate measured evidence from claims, for example with no invented growth, observational-not-causal outcomes and documented catalog provenance, which builds real trust.
To improve
Reconcile the inconsistent numbers in your materials. Compute is quoted at both $0.012/hour and $0.252/hour, and the 22-second call is reported as taking both 10.2 seconds and 205.2 seconds, with stage timings that do not add up. Your 94.5% precision also needs its calculation explained, given one extra finding out of eight. Your benchmark is eight synthetic cases scored by lexical matching, so a small human-labelled set of realistic calls would strengthen the accuracy claim considerably. Run even a rough timed manual-review baseline and estimate cost per processed audio hour, since these are the adoption metrics your track asks for.

The team

Kamran Mammadli
developer
Agil AghamirzayevLead
developer
Bakir Gahramanov
data
Hasan Nurmammadov
data

Disclosure

Required by the rules. Judges verify it against the repository.

Models: Public pretrained models: Systran/faster-whisper-small for transcription; SpeechBrain spkrec-ecapa-voxceleb for mono speaker separation (stereo channel metadata maps roles); Qwen3 4B Instruct via Ollama for structured findings, reports and entity masking; nomic-embed-text for issue similarity. No model was trained or fine-tuned by the team. Optional GLiNER support exists in the code but is not active in the deployed demo. Rule-based masking and evidence validation supplement model output.
Data: Eight fictional, generated support recordings with local transcripts and AI outputs; eight additional synthetic text evaluation cases and short synthetic stereo benchmark calls. No private customer recordings or proprietary training dataset is included. Prepared catalog outputs are stored older model results with editorial titles and documented masking corrections (demo/README.md); opening them is not fresh inference. New uploads use the current pipeline. Synthetic labels and keyword matching are development evidence, not independent human-reviewed accuracy. Raw audio itself is not redacted.
Components: Reusable open-source components: React, TypeScript, Vite, Radix UI, TanStack Query/Table, React Router, Recharts, Spring Boot/Security/JDBC, Flyway, PostgreSQL/pgvector, MinIO, Kafka, Temporal, Ollama, Python, Pydantic, faster-whisper, SpeechBrain, scikit-learn and Caddy. Versions are recorded in lockfiles, Maven configuration and container definitions. Custom work includes the review workspace, tenant isolation, recording ingestion, versioned company instructions, evidence validation, issue grouping, human decision lifecycle and evaluation harness. Pretrained model weights and infrastructure components are reused, not claimed as original.
Built during the event: Confirmed

Quality testing

VALIDATION: 131 automated tests passed across five layers: 9 frontend unit tests, 58 worker tests with database and real local models enabled, 27 API tests, 34 browser tests and 3 live guest browser tests. Production build and lint checks also passed. Commands, logs and the scope of each recorded run are documented in docs/SUBMISSION_EVIDENCE.md. DEPLOYED ACCEPTANCE: We tested the complete recording-to-insight flow on Azure. A fresh 22-second upload produced nine transcript segments and one accepted issue in 10.2 seconds: transcription and masking took 9.5 seconds, AI analysis 12.6 seconds and clustering 0.7 seconds. A separate report across nine calls generated three recommendations in 8.7 seconds. HTTPS, timestamp-linked audio playback and guest isolation passed. Cross-visitor access returned 404; requests without CSRF protection returned 403. CONTROLLED COMPARISON: Clarity and a frozen keyword baseline were evaluated on the same eight new synthetic transcript cases. Clarity matched 7/7 labeled issues; the baseline matched 3/7. Automated label-matching recall increased from 42.9% to 100%, a 57.1 percentage-point improvement. Clarity achieved 94.5% precision versus the baseline’s 100%, producing one extra finding. Three additional rubric flags remain documented for review. These results measure lexical matching on a small synthetic set; independent semantic assessment remains pending. FAILURE-DRIVEN IMPROVEMENT: Testing exposed cases where invalid evidence could interfere with otherwise valid findings. Version 6 validates findings individually, preserves accepted findings and rejects invalid or agent-only evidence. Uncertain speaker roles trigger source excerpts and visible review notices. A keyboard-focus defect was corrected before all 34 fixture browser tests passed. Deployment failures also led to a corrected web image and a reproducible build of the required MinIO release from official source. NEXT QUALITY GATE: The repository includes failure cases and a prepared reviewer study comparing manual review, keyword assistance and Clarity. It will measure correctness, unsupported recommendations and active review time. Representative noisy-audio and privacy testing, sustained-load testing and cost per processed audio hour remain explicit validation milestones.

Worth 20 of the 100 points.

Running it

LIVE DEMO: Open the demo URL in a normal browser. No account, installation or AI API key is required. Each visitor receives an isolated temporary workspace with eight fictional recordings and prepared evidence. FRESH AI SCENARIO: Choose Try fresh analysis -> Conversations -> Analyze sample call. While it runs, inspect the prepared conversations, issue evidence and Decision center. When Ready to review appears, open the new call, play its audio and inspect the transcript and linked issue. In Support summary, Refresh analysis generates a new report. On our Azure CPU acceptance run, the 22-second call took 205.2 seconds; the nine-call report took 241.7 seconds. These are individual measurements, not latency guarantees. DEPLOYMENT: One Azure East US VM, 4 vCPUs, 32 GiB RAM, 128 GiB SSD. React, Spring Boot, Python worker, PostgreSQL/pgvector, MinIO, Kafka, Temporal, Ollama and Caddy run together. HTTPS is enabled. Inference runs on the VM. Compute is quoted at $0.252/hour (about $6.05/day), with disk and public IP billed separately. Trial credit funds the demo. REPRODUCE: README.md links the complete Docker Compose setup, generated private credentials, model provisioning, test commands and machine-readable Azure acceptance evidence. Public guest integrations are simulated and send no external messages. Five uploads per guest, 25 MB/five minutes per file, 24-hour sessions. Next step: a consented pilot with independent labels, manual-review timing and measured cost per audio hour.

Project ownership

The team keeps the rights to this project. Judges were given access for verification only.